Sean McMillan: Super Smash Bros. Melee is nearly 25 years old, but it's still one of the most competitive fighting games in the world. And wherever you find competition, you're eventually going to find people looking for an edge. Our next guest is Dan Petro, along with maintaining the SLP Enforcer tool. He helps investigate cheating allegations within the melee community. Today, we're talking about the surprisingly sophisticated ways people try to cheat and how they get caught. Dan, thanks for joining us.
Dan Petro: You're welcome.
Sean McMillan: ⁓ so when I first saw this come across my desk, you know, I get notifications about when people are speaking at conferences about this and that, cloud hacking, about AI development, all these kind of things. And when something comes across about Super Smash Bros. melee, it kind of sticks out. ⁓ I I guess my first thought of how you cheat in melee.
Dan Petro: Mm-hmm. Yeah.
Sean McMillan: Because we ⁓ any of us who have played it accuse our friends sitting on the couch right next to us of somehow cheating. I picture like a game genie or something tangible like that. ⁓ but what's it what's it actually look like?
Dan Petro: Mm-hmm. Right. Yeah. Yeah, that's a good question. So first off, ⁓ the GameCube is a disk based system, so you can't use a game genie. ⁓ that wouldn't work. But ⁓ yeah, no the point stands. So ⁓ yeah, so the ⁓ Super Smash Bros. Melee for the Nintendo GameCube, as you said, is ⁓ just about twenty five years old. ⁓ I was there the day that it came out, I've been playing it ever since. ⁓ so this has been a hobby of mine for a really long time. And ⁓ ever since I've been into information security, ⁓ I've you know taken this ⁓ hobby in a different direction. I've been trying to
Sean McMillan: Yeah.
Dan Petro: ⁓ like automate playing the game and ⁓ lately it's been around like trying to investigate ⁓ cheating ⁓ inst instances. So what does that look like broadly? ⁓ I I think that one good way to think about it is to split it up into two distinct categories, ⁓ offline and online. So the GameCube of course ⁓ is in the offline era. It doesn't even have internet capability. If you wanted to get your GameCube online, there isn't even a way to do that. Within asterisk ⁓ I think technically there
Sean McMillan: If you've listened to this show for a while, you've probably noticed we keep coming back to trust. And I swear we're not doing it on purpose. It just keeps showing up all over the news. So every week when we're talking about AI, cloud security, identity, incident response, any of that stuff, it always comes back to this question of what assumptions were people trusting that turned out to not be true. And this week, that question kind of takes an interesting turn. There was the little attachment for Fantasy Star Online. I remember that. Yeah, yeah. Yeah.
Dan Petro: is like th technically there's an attachment that like only existed in Japan like for fantasy star online but like 99.9% of the time there there isn't a way to ⁓ there there's no online games on the Nintendo GameCube. So like what are you talking about an online mode? So you have to understand that ⁓ for melee specifically, ⁓ there is a community created and run online mode ⁓ called Slippy. ⁓ you kind of see the green thing right here. This is for ⁓ Project Slippy ⁓ that ⁓ added a replay system into the game and then also kind of built into that
Sean McMillan: Because it's not just us asking it, it's the people building the technology. So this week we'll look at an AI agent that inherited authority it was never meant to have, how AI is reshaping bug bounty programs, another major breach at Origin Energy, why employees at OpenAI, Anthropic, and other Frontier Labs are publicly asking governments to help slow down AI development.
Dan Petro: ⁓ online mode. So you can play against opponents using rollback net code ⁓ seamlessly online. ⁓ it's great. And this happened a Right in the middle of COVID too. So he was like right in at the time when the community couldn't meet up in person, but we were finally able to play this game online in like a really good way. That ⁓ puts a lot of ⁓ online modes that major other publishers ⁓ created to shame, too. Like the online mode is really good. ⁓ so ⁓ the offline ⁓ instances of cheating are like ⁓ maybe a lot more about what you might have expected. ⁓ so actually, before I get into that, I wanted to
Sean McMillan: And I'll sit down with Dan Petro to talk about catching cheaters in Super Smash Bros. Melee, where trust, verification, and adversarial thinking are just as important as they are in cybersecurity. This is Initial Access. I'm Sean McMillan, community manager here at Bishop Fox, and I'm joined today by Sergio Viegas, Senior Managing Analyst, and Kendrick Urbaniak, Senior Operator. Gentlemen, thanks for stopping by.
Dan Petro: Mention one thing. The word cheating ⁓ is very loaded. ⁓ I come also from the background of Magic the Gathering, as you can probably tell from behind me, ⁓ which has very specific ⁓ definition of cheating that I actually really like. In order to be cheating, ⁓ it has to be three things. It has to ⁓ you have to have one violated the rules. ⁓ so that's obvious. But that's not only one category, that's not all three. You needed to have done so on purpose.
Sean McMillan: Yeah.
Kendrick: Good to be.
Sergio Villegas: Always a pleasure.
Sean McMillan: ⁓ as always, before we get into the headlines, I want to do a quick look at what's coming up around Bishop Fox. And ⁓ boy, is it a lot. If you're heading to Black Hat, come say hi. I'll be at booth 5200 and I'll be there throughout the week meeting folks, recording interviews for the show. ⁓ so if you're around and you've always wanted to tell me how wrong I am about something, this is your big chance.
Dan Petro: And you needed to do so in order to gain a competitive advantage. ⁓ So the all three things must be true in order for it to be cheating. Otherwise, it's just an accidental mistake or something else happened. So some other good examples, I I can give some good examples as we go forward as to like, you know, why ⁓ all three things are really important.
Sean McMillan: Mm-hmm. ⁓ then over at DEF CON, we've got a packed schedule with talks across several villages, including AI Village, Cloud Village, Red Team Village, Game Hacking Village, where Dan Petro will be presenting, ⁓ Noob Village, and OSINT for good. If you're making the rounds, you're probably gonna run into a fox. And if you're not going to Vegas, we've got a couple virtual sessions coming up on August 4th. Join the Cosmos team here. Yeah.
Dan Petro: So ⁓ a lot of the cheating that you'll see ⁓ that might occur offline, meaning at real in-person tournaments where you're sitting across from somebody, right? It's a a high-stakes environment. People care about these tournaments. ⁓ but also ⁓ it's an environment that is inherently difficult ⁓ to cheat at because you're sitting right next to your opponent. So like if they plugged in a computer like to the GameCube or something, right? Like you might notice. ⁓ shenanigans are much more physically apparent.
Sean McMillan: For the prioritization problem, why more findings don't mean less risk, a discussion on continuous threat exposure management, cutting through AI-generated noise, and focusing on what attackers can actually exploit. And that'll have our own Richard Brown, friend of the show, on that. Then on August 25th, we've got another virtual session coming your way. Yeah, yeah.
Dan Petro: ⁓ when you're in that kind of environment. But nonetheless, this is a problem, specifically because of these things. This is the GameCube controller. ⁓ As I mentioned, ⁓ the GameCube is twenty-five years old and they don't make these things anymore. And so if you wanted to play Melee
Sean McMillan: Mm-hmm. ⁓ And that is gonna be where AI breaks field notes from Gen AI and agentic testing. So, with that out of the way, let's get on to our top stories. First up this week from the Hacker News: Chat GPT agent forger flaw could deploy rogue workspace agents via a phishing link. Zenity Labs disclosed.
Dan Petro: ⁓ you need a good controller, ⁓ one that can ⁓ like play the game at a high level and the analog sticks in particular, ⁓ this over here, wears out over time. And ⁓ because they don't make these things anymore, that means that the price started to skyrocket. And so one angle pi ⁓ the communities come at this from a bunch of different angles. ⁓ one of them is that we've created a number of they call box style controllers. So if you're more familiar with like a Street Fighter or something like that, that's like a fight stick where
Sean McMillan: Sure. Agent Forger, a flaw in OpenAI's ChatGPT agent builder where a URL parameter meant to pre-fill a prompt box was instead automatically submitted and executed the moment the page loaded. A single crafted link opened with no confirmation dialogue could silently create, authorize, publish, and schedule an autonomous AI agent inside that employee's workspace.
Dan Petro: rather than like analog sticks, it's all just buttons. ⁓ And so they made a number of these ⁓ controllers that we can create as a community, basically. ⁓ And there's arguably some ergonomic benefits to basically playing the game, typing like a keyboard, but although all that aside, certainly there's a strong argument just from the perspective of like, well, it's a community created controller.
Sean McMillan: Mm-hmm. The forged agent inherited every enterprise connector already approved. So email, calendar, cloud storage, Slack, Teams, any of that stuff without any new consent to alert anyone. Once live, it pulled the attacker's inbox for commands on a five-minute schedule, turning ChatGPT into a covert C2 channel that could exfiltrate data and impersonate the employee indefinitely. Yeah. ⁓
Dan Petro: ⁓ so a lot of people use these things, they want them to be ⁓ in place. but ⁓ it also is a strong rift ⁓ since ⁓ now you're controlling your character in a completely different way. And so long story short, ⁓ the community created a rule set ⁓ that says like those style controllers that are fully digital that don't have any analog ⁓ pretend ⁓ buttons on them ⁓ have to adhere to a certain set of rules. ⁓ so like when you move the ⁓ when you press buttons to move
Sean McMillan: Okay. OpenAI fixed the issue ⁓ June 8th. Zenity's Mike Takahashi frames it as an agent trust failure at its core. And the platform trusted that the user had intentionally created and operated the agent when none of that was verified. So ⁓ this is this is always my fear with these agent kind of situations where if it's really good, no one's gonna detect it because it's working exactly like it should.
Dan Petro: Your joystick like left and right. It has to have an emulated amount of travel time between. Otherwise, if you just press left and right, it'll instantly switch between left and right. And there's reasons why that can be very strong in the game. ⁓ Or there's ⁓ restrictions around ⁓ like your dash dance. There's some restrictions specifically around Smash DI, which is like this silently broken technique inside of melee that humans are not able to fully exploit just because it's incredibly difficult to perform. But it's actually a lot easier.
Sean McMillan: Mm-hmm. Yeah. Dash dancing, right? Yeah. Yeah. ⁓ but but what does it kind of reveal about these platforms when the thing that's trusted is a URL parameter rather than any actual like bizarre user action?
Dan Petro: easier on those ⁓ box style controllers. ⁓ and so there's some rules around like the basically at the firmware level on those ⁓ controllers they have to have restrictions on how fast you're able to like move your joystick back and forth, things like that. To nerf the controller in some way.
Kendrick: I I think i this is is a scary thing. ⁓ you've heard of like regular cross site request forgeries. ⁓ now we have that with AI. Let's go make some AI agents in people's workspaces. That's pretty ⁓ that's pretty bad, right? ⁓ and persistence game is could probably be pretty ⁓ well ⁓ fleshed out, probably, ⁓ to the point where
Sean McMillan: Yeah.
Dan Petro: ⁓ and so the first and foremost way to ⁓ to cheat, quote unquote, at a ⁓ tournament in Super Smash Brothers melee is to use one of these controllers that is illegal, that the community has deemed to be ⁓ not legal. ⁓ and so ⁓ this is a problem that can be somewhat automated. so I created a tool, ⁓ it's called the ⁓ SLP Replay Enforcer. so after every game of melee that is played.
Sean McMillan: Yeah.
Kendrick: AI, it doesn't you can probably have some code spin up your AI agents and take them down because right now I think workspace agents are kind of visible for the most part. So I feel like from the most part you would see this to some degree, at least running, and you would ask yourself why is it running? So I don't know. I feel like there's definitely some more tricky things that we're gonna have to start looking out for in the future where if this is actually, you know, taken to the next little step. ⁓ I know this was like a research paper-ish type thing.
Sean McMillan: Mm-hmm.
Dan Petro: on a in a modern setting, both online and at a major offline tournament. The game using Slippy will record a replay of the game. So a replay is not a movie file, it's not just like a video of the game. It basically records all the button presses that both players had during the game. And then also some game state information about like whether you're standing or not or what exact animation the character is in, things like what your current health is, things like that.
Kendrick: So ⁓ I definitely think we're gonna have to see where this goes because I think it can get much much more dangerous from here.
Sean McMillan: Yeah.
Sergio Villegas: Hundred percent. And and I think you you mentioned it, right? Like the the key in here is that it worked as intended, right? I was doing some some research on on this specific topic. It was it caught my eye, right? Like ⁓ even if you set your own like your own desktop application in like ⁓ supervised mode where you have to ⁓ approve everything, it doesn't matter because
Sean McMillan: Sure.
Dan Petro: And so what we do is ⁓ we can run those replays through a like statistical analysis checker, also some ⁓ hard heuristics to say like what is the likelihood that this was played using an illegal controller? Like what is the like ⁓ what what are the odds of that basically? ⁓ so in some cases it's a statistical analysis, and we give a direct odds breakdown, like ⁓ are the chances of like this ⁓ having been performed at random.
Sergio Villegas: It's all within the agent parameters and trust that you gave to the agent. So it doesn't matter if you have not given in permissions to do all of that. The agent by itself, by its nature of the agent alone, it's it's working as expected, right? And the problem with that is how much like, yeah, when when I saw like OpenAI fixing the issue, like should have they, right? Like it's it's not a problem. That's how the agent was configured, right?
Sean McMillan: Yeah.
Dan Petro: versus like you know ⁓ using a legal controller. In many cases we can just say specifically at specific ⁓ heuristic based evidence that like this is an illegal series of inputs. Like you're not allowed to press the buttons this fast basically. And so that way ⁓ tournaments can automatically plug replays into this tool. It's already built into you a lot of the ⁓ systems that tournament organizers use ⁓ to automatically detect like ⁓ this person's using a controller that is not kosher. ⁓
Sergio Villegas: It's the pr it they they fix it in a way of like putting more gat guardrails, right? ⁓ but I think it's a problem of one awareness, right? Like how much are we people trusting our agents and what random agents we are we importing into or ⁓ I I know we throw ⁓ every every episode we throw shade at AI, but in this specific case it's ⁓ more like to the humans using the AI, right? Like hey, what are you importing?
Sean McMillan: Yeah. Yeah.
Dan Petro: And additionally, it's not just those box style controllers, even these guys ⁓ can hide small ⁓ computers. ⁓ so the often they'll have like an Arduino built into the case. And you can use that to automate in some cases like full like button presses. You can run macros. So like if you just like surreptitiously press like a button down here, it'll like perform some combo automatically for you or something like that. ⁓ and so we can do some checks to like look for that kind of thing too.
Sean McMillan: Ex that's what I was gonna say. It's not shade of AI, it's shada poorly using AI. and agentic stuff is like it's agentic, it's going to take action. So like you gotta think about that before you enable something.
Sergio Villegas: Yeah.
Kendrick: Yeah. Well I I mean
Sean McMillan: Mm, sure. ⁓
Kendrick: You you can't blame the user one hundred percent on this one, 'cause, you know, they just visited a URL. And based on reading the the paper, my understanding is that just by visiting it executed the the the hook to to create an AI agent. So you can't necessarily blame the user on this one a hundred percent. It was definitely a bug that needed to be fixed, and I'm glad they fixed it. But I think we're gonna have to watch out for
Sean McMillan: it seems like the sky's the limit with what you could do inside one of those controllers. Are there any cases where ⁓ you just kinda had to stop and All right, that's that's good game. ⁓ That's a that's a clever hack. ⁓ Sure. Yeah.
Dan Petro: Huh. Yeah, there ⁓ well so at least there are many instances of the rule set. So there is this ⁓ document you can ⁓ you look up that is the official controller rule set. And many of the restrictions inside of that are ⁓ just empirically unenforceable from a looking at the button progresses perspective. So one really good example of this is that ⁓ there's a restriction on the location of the buttons on the face of the controller. So that
Kendrick: now's the time to take stock of your own workspaces and making sure that you know all the agents that are in your own workspaces and in the future keep track of your workspaces to some degree and you know audit them. So yeah.
Sean McMillan: Exactly. I think more than anything it's a wake up call. Like, yeah, I d I don't think it is fair to to blame the user in this case, but it is like, holy smokes, like this is gonna make me think twice about how I operate in the future, you know? ⁓ and I think a lot a lot of people a lot of people should.
Dan Petro: you're basically you're not allowed to put buttons on the backside of the controller. ⁓ and there's reasons for this. ⁓ that
Sean McMillan: Sure.
Kendrick: Also, maybe stop running AI agents on your phone where you're visiting more random links in general. So yeah. Don't
Dan Petro: We don't have to get into here. ⁓ But ⁓ if you're looking at a digital file of button presses, there's no way to tell where the button physically was, like was it on the back side of the there's just no way to do that. ⁓ so basically not every ⁓ rule is enforceable from this perspective. But that's one of those things where when you're there in person with someone at an offline event, like you could just see there's a button, like they're pressing a button on the back side. ⁓ so it's ⁓ pretty safe ⁓ from that perspective.
Sean McMillan: Yeah.
Sergio Villegas: Right.
Sean McMillan: Mm-hmm. Yeah. Yeah. We're not checking everything on our phone. We're we're just not.
Sergio Villegas: Yeah. ⁓ talking about like next steps, something I'm envisioning is like I know with cloud you have this remote mode. And I I was thinking exactly what Kendrick mentioned, right? Like if you have remote mode enabling your phone and your desktop and you got that cross request on your phone, it gets sent to your desktop. And you know how much more access you will have in your like it will be kind of convolated, ⁓ but it's feasible, right? It's feasible that and also ⁓ something that I've seen at a trend is
Sean McMillan: Yeah. Yeah.
Dan Petro: ⁓ things do get hairier when you talk about online events. so online events are ⁓ speaking treated less seriously with less ⁓ gravitas ⁓ ⁓ offline events do. ⁓ that said we still do care about them, and so ⁓ we tried as best as possible to prevent ⁓ whatever shenanigans we can ⁓ the online space as well. And so this is ⁓ it encompasses the offline instances as well. So that ⁓
Sean McMillan: Mm-hmm. Yeah.
Sergio Villegas: these ⁓ these appli these ⁓ add-ons for your Chrome, your Firefox that uses your ⁓ your AI account to scrap to scrape the websites. So it's kind of same situation, right? Like just by visiting the site and just because you have that add-on enabled by default, it will scrape everything and run things, right? So ⁓ yes, maybe don't blame the user, but at the same time it's like just remember that you're accessing random websites. Maybe I it's it's it's tricky. I know it's tricky.
Sean McMillan: And
Dan Petro: the the same ways you would cheat offline by using illegal controllers, things like that, ⁓ apply here. But then also there's some unique online specific considerations. So one good example of this was ⁓ somebody who just tried to ⁓ fake lag. So ⁓ the game tries to adhere to ⁓ tries to use ⁓ its rollback netcode system to ⁓
Sean McMillan: Yeah. That's why I'm still using a Mac two over here. Like that's why my camera cuts out all the time.
Kendrick: Things ⁓
Sergio Villegas: Yeah.
Kendrick: This is a small callback. Th this is a small callback to what I said ⁓ a couple of episodes ago maybe, ⁓ where I said ⁓ websites in embedding anti-agent ⁓ stuff in them. Well, now you're probably gonna start seeing, you know, that used maliciously to some degree as well. So ⁓ we're we're seeing a natural progression, I feel like.
Sean McMillan: Mm.
Dan Petro: smooth over a lot of the rough edges of network latency in an online environment. But it's not always perfect. And so sometimes if you've got like a really bad connection from person to person because the game is peer-to-peer, ⁓ then you might be having a suboptimal game experience basically. And so there was definitely ⁓ one instance in which a person had ⁓ played an official game ⁓ that was in an online tournament and then afterward tried to claim that like the game was super laggy and really bad for them.
Sean McMillan: Yeah. All right. So let's see. Also from the hacker news this week. ⁓ GitHub cuts bug bounty payouts, ⁓ moves top to VIP tier. ⁓ file under had to happen sooner or later. Beginning July twenty seventh. GitHub cut its public bug bounty payouts by at least half at every severity tier. Yeah.
Dan Petro: at that ⁓ they wanted to have the game be replayed. and so this was a thing that got brought to my attention after the fact. This was all post hoc. ⁓ and then ⁓ asked, like, was this real? Like, was this actually a laggy game? How could we possibly know? And so ⁓ again, the replays can tell the tale here. since ⁓ the way that the input system works inside of a rollback netcode system is that ⁓ anytime that there is a network lag event, it gets reported as a rollback. So ⁓ we see that in the inputs.
Sean McMillan: Critical findings drop from as much as 30,000 to a flat 10,000, while creating a permanent invite-only VIP tier that keeps critical payouts starting at 30,000 for vetted researchers. GitHub says the restructuring is meant to cut through a flood of low-quality AI-generated reports, overwhelming its triage team. And it's the fourth major bug bounty program to restructure or narrow its scope in 2026, joining Google, Bug Crowd. Yeah.
Dan Petro: ⁓ so we can basically see all the instances in which ⁓ rollbacks happened, and that will be a metric that can tell us pretty directly one-to-one how laggy the game was. It's right there in the replay file. So it was simply a matter of gathering the replays from both players and then seeing from both of their perspectives, because of course it's ⁓ peer-to-peer game, ⁓ like was it a very laggy game? And we could just see that the answer was no. ⁓
Sean McMillan: And Hacker One's Internet Bug Bounty Program. Newcomers to this public program now face a four submission trial period before they can build the track record needed for VIP consideration. You don't earn more by submitting more, GitHub said. You earn more by submitting better. GitHub coming in with the dad advice. I love it. I I would love to have that power when I play Rocket League and my teammates say, I'm sorry, it's lag. No, it's not. ⁓ the change follows a May 2026 policy shift that already required working proof of concept exploits and demonstrated impact before submission. what does this do to to the incentive game? I mean, does this does this take the wind out of anyone's sails or is this kind of just expected?
Sergio Villegas: Mm.
Dan Petro: Yeah. Yeah.
Sean McMillan: what's the what's the process look like? If someone reports ⁓ a suspected cheater, ⁓ what's an ⁓ investigation start as? ⁓
Dan Petro: Yeah. Yeah, so again, ⁓ there's a pretty good split between on and offline. So at an offline event, there's typically a person who is there at the events. ⁓ so at some of the larger events like Genesis that just happened, it was ⁓ a good friend of mine, is Eric Lee, goes by violence ⁓ online. And ⁓ so he's there like in person checking controllers running them through the tool that I created, and then we're often in direct conversation about like, you know.
Kendrick: We already s we're we already saw this though. Like w there when Bug Bunny first started, it was a lot of public programs. And then a lot of programs started going private because they were getting a lot of submissions. And now we're seeing AI, and then even those private ones are probably not too happy with the amount of AI that's being used in them. So I I think we're just again seeing the the the ramifications of everybody having AI and everybody using it, and then all these companies not being happy that
Dan Petro: But is this suspicious or not? So, one of the troubles, of course, with this whole scenario, and why I wanted to circle back to the three rules of whether or not something is cheating, is specifically this problem: that just because you ⁓ a controller that's not technically legal doesn't mean you're cheating. So ⁓ this in a few varieties. One is which one of which is that if you just go on to Amazon or whatever store it is and buy one of these rectangle controllers, ⁓ it will
Kendrick: Getting all these low quality reports, and then we've seen ⁓ lots of different programs either cut the bounties entirely to incentivize people to stop looking, or maybe only look at the people that are actually dedicated to securing those particular products because those are the only people that are gonna submit to no bounty programs for the most part. ⁓ and now we have these higher-end programs that are generally they have public portions of them, but most of the time they are they're running private programs as well, who are also now.
Sean McMillan: Mm-hmm.
Dan Petro: almost surely come to you with a firmware that is not considered legal by the community. Which is unfortunate. That's not great. We don't like it. We've been trying to work with the vendors there, but that's just simply the reality.
Kendrick: cutting their payments because it's just getting too much. And and now you're gonna have that one person that submits really like three good high quality reports and on their third one or their fourth one, they're gonna some more submit that really low quality one and then they're gonna get banned from the program and then we're gonna see a non-responsible disclosure and because they're disgruntled. So who knows? we're probably gonna see some some wild west ⁓ stuff going on here soon. So
Sean McMillan: Mm-hmm.
Dan Petro: and so if somebody just buys a controller and just doesn't know all this stuff and then shows up to their like local or shows up to even a like medium or large size tournament, they might be violating the rules and literally not know about it. And they're not trying to do it to cheat, they just like weren't aware that this whole thing existed. And so at the tournament, you have the ability to ⁓ like reflash the firmware of the ⁓ controllers basically. So once you're ⁓ at least at Genesis, the ⁓ like large, the probably the biggest tournament of the year.
Sean McMillan: Yeah. Yeah.
Sergio Villegas: Yeah. And and we have seen this happen before, right? With companies cutting down their PDPs and something that almost always happened, right? Not saying it will happen, but will likely happen is ⁓ Yeah, yeah. ⁓ looking at one I I think Kendrick mentioned right. One is the very private ⁓ companies that buy your exploits, your zero days, right, for themselves, for their testing.
Sean McMillan: No.
Dan Petro: But the way they ran it was round one, they just let everybody play. They we then investigated all the inputs from all the controllers for everybody ⁓ played from round one, found the people that were using the illegal controllers, and then did a follow-up on those. And the vast, vast majority of them are just like people who, you know, found their way into a borrowing a friend's controller and they're not a serious competitor, they're just like borrowed a controller or they just bought this thing online and they just used it the first time or whatever. ⁓
Sean McMillan: Crossing fingers here. Yeah.
Sergio Villegas: And the other thing is the the dark net markets, right? Like dark dark net dark web markets. Where it's very competitive where you are getting tractors paying ten, twenty thousand for a zero day for an access, right? So when ⁓ it's it's hard for operators, right? Like let's think about a ⁓ an operator, ⁓ book bounty person that does that the entire year, three year round, right? Like he can get either top ten thousand from ⁓
Sean McMillan: Yeah.
Dan Petro: Or in in some cases, they're clearly not using it for competitive advantage, where you have some people who try to do like fun things, like they'll play with a guitar hero controller, because you could actually do that. So you can you can like get an adapter and play the game with guitar hero, like you're like strumming to like dash dance or something. ⁓ they're not gonna be in the top eight, they're not, you know, they're just memeing. Just let them have their fun. It's technically illegal, but like whatever, right? They're not doing this for competitive advantage.
Sean McMillan: Amazing.
Sergio Villegas: From GitHub and and that's it. And he needs to pass the the bet the whole Betoin system to get approved and whatnot, right? Or he can get twenty thousand from a random state sponsor actor and call it a day, right? I know it it's it's it's very intricate, right? Like I know ethics and whatnot, but let's be realistic, right? Like it's their job, they want to get money. Again, hopefully they go the right way, but this has happened before. Yeah.
Sean McMillan: Yeah. Let ⁓ have their fun. Yeah. Yeah.
Dan Petro: ⁓ and then every once in a while you run into somebody who is like a top 100 player. So who's a known quantity, the community knows them, like their name would be somebody you would recognize, and they show up round one using an illegal controller, you're like, ooh, like that is more suspicious. ⁓ like that's somebody who should know better, ⁓ who nonetheless shows up to the largest tournament of the year using a firmware that's not legal. ⁓ so that's the kind of thing that we are like really keeping our eyes out for.
Sean McMillan: ⁓ Not everyone's going to. Yeah. Bad luck. Yeah.
Kendrick: Yeah. Well and the the barrier to entry is getting substantially lower for finding the basic vulnerabilities now. Like AI can do it, right, for the most part. ⁓ sometimes it refuses and you gotta poke and prod it, but you know, it'll do it eventually. ⁓ so ⁓ that is probably another thing where we're gonna see is is that you're right, Sergio, we're probably gonna see maybe a little bit of a of a rebound because
Sean McMillan: Yeah. No, do you ⁓ do you find ⁓ and you don't have to name names, but you certainly can. ⁓
Sergio Villegas: Mm-hmm.
Dan Petro: Yeah. Also in the in the ⁓ in the DEF CON talk, I'm not gonna be naming specific names. So we're not gonna be dropping any specific accusations against sp the specific people. These have all been taken care of, so it's gonna be fully anonymized. ⁓ so FYI.
Sean McMillan: That's right.
Kendrick: to the to the darkness slash zerodium, whatever whatever they call themselves these days, ⁓ that's, you know, buy and sell exploits. and it's just like, you know, what can you do? Like these people probably want money and they used to have a legitimate way of getting it and now they're
Sean McMillan: Right. no, do you ever find like people who they're they're detected, they're caught basically. And whether they were intentional or not, ⁓ off the off the bat, or you thought it was intentional or not, do you see people kind of repeat offending and trying to get around your detection and like kind of the the cat and mouse game of cybersecurity?
Sergio Villegas: Okay.
Dan Petro: Mm-hmm.
Kendrick: gonna turn to probably less legitimate ways of of getting it, which is very unfortunate. I do wish these bug runny programs would maybe take another look ⁓ at how they're in taking ⁓ vulnerability reports, ⁓ fight AI with AI, ⁓ just like ⁓ a
Dan Petro: Yeah. Yeah, ⁓ I do think that w we don't see a ton of ⁓ like cat and mouse stuff specifically in melee. There I think that th I think there's maybe there's a reason th a couple of reasons for that. one is that ⁓ the effect of your controller on the game is pretty subtle. ⁓ one of the the underlying realities behind all this is that ⁓ melee is a game
Kendrick: places have done. ⁓ maybe that's better solution so you don't alienate people and make it harder for people to report vulnerabilities to you, especially when they're in the critical range of, you know, we're only gonna pay you ten thousand dollars for this critical because you're not part of our trusted program. It's like, ⁓ okay, yeah. So yeah, I think you're right, Sergio.
Sean McMillan: Yeah.
Sergio Villegas: Yeah. Yeah. Maybe we can survey with with our maybe we can survey with our ⁓ with our audience, right? Like hey, do you want to get paid that ten thousand or twenty thousand? Yeah, yeah.
Sean McMillan: Yeah, platinum status.
Dan Petro: Melee as a game is extremely low variance, where the better player wins almost every time. It's extremely high skill cap. a great example of this is that a very good friend of mine who I ⁓ played a bunch with in high school as a go ⁓ I I played back in the day was ⁓ Tony as a Taj in the local ⁓ Arizona melee scene. ⁓ And I have, and this is true, never taken a single game from him ever. And we must have played ⁓ many of games. ⁓
Kendrick: I well I think most people can do math.
Sean McMillan: I think I know what they're gonna say. Mm-hmm. ⁓ Yeah, I you know, I almost feel worst for the ⁓ influencers because I know they're out there. I don't know how I haven't been fed this based on my social media usage. But like there's gotta be influencers out there that are just saying, like, I made thirty thousand dollars this month without even doing anything because my agents are doing bug bounties. Those those guys have a rude awakening. That's I'm so sorry. Yeah.
Dan Petro: just hanging out, like late at night at a friend's place where he's playing Bowser and I'm trying to beat him. ⁓ like it it it doesn't matter. ⁓ three AM in tournament, I've literally never beaten him ever. he's just an order of magnitude better than me at the game and nothing I ever did mattered. and so ⁓
Kendrick: Th those people exist, yeah. Well, th it's not gonna stop them for the most part, right? They'll probably just refine their reports and maybe add an extra agent to verify their reports. Like it's gonna cost them maybe an extra grand in time and cloud usage or whatever agent they're using. ⁓ so I mean at the of the day, like I hope they take the better thing of like, let's improve their own bug bounty reporting system rather than, you know, going to the relative dark side of of vulnerabilities.
Dan Petro: It doesn't matter what controller I was holding, it wasn't gonna make a difference in that. ⁓ it it does make a difference at the very top level where like you're trying to eke out tiny advantages. ⁓ but for the very vast majority of games, ⁓ what controller you're using kind of just doesn't matter.
Sean McMillan: Yeah. Sure. Yeah.
Sergio Villegas: Yeah.
Sean McMillan: Sure. All right. Let's talk for a second about ⁓ this from Security Week 2. Origin Energy. Australian energy giant Origin Energy confirmed a breach after a hacker claimed on July 24th to have stolen the data of two million customers and threatened to leak it. At the time of reporting, Origin had not disclosed the initial access vector or independently confirmed the attacker's claimed numbers. Yeah.
Dan Petro: ⁓ so there's that. ⁓ that th just just take that as background information ⁓ where like that's kind of why we mostly care about like what the top players are doing, like basically the top 100, ⁓ because those are a the most likely people who are like th if you're going to try to cheat those are the people who are going to do it. ⁓ those are the people for whom it would matter the most. kind of this like low level like ⁓ like somebody who's you know gonna be making the top five hundred of a tournament.
Sean McMillan: This marks Origin's second major disclosed incident within months. Company already reported a separate breach affecting 900,000 Australians earlier this year. Origin is a major electricity and gas retailer, ⁓ putting billing, contact, and account data for millions of Australian customers at stake across both incidents, and no ransomware or extortion group has yet been publicly tied to this July claim. What does it usually mean?
Dan Petro: Like they're just not going to significantly improve their standing by using an illegal box controller. ⁓ that's just the mechanics of the game. ⁓ so I think for that reason we haven't seen a huge influx of like cat and mouse, of like people trying to push the limit of like what you can get away with ⁓ in a ⁓ a controller. ⁓ and there's the offline aspect of it where like
Sean McMillan: Yeah. I think. ⁓ ⁓ for from an operator, when a second major breach at the same company surfaces within just a few months of the first, like, are we thinking same door left open or is this just like Swiss cheese, just holes everywhere?
Dan Petro: ⁓ you're sitting there next to the person, so the amount that you can get away with is kind of limited in terms of like macros or like other funny business going on with your controller. online things are ⁓ different. Where ⁓ it is a small and insular community, ⁓ and so often even in an online setting you know the person who you're playing against. ⁓ there there isn't many millions of people who play this game.
Sean McMillan: Yeah.
Kendrick: I think from the operator side of things this is more of a Swiss cheese thing that is going on. Or they're just really good with their fishing and they just got like, you know, two people relatively same time-ish. ⁓ they got a really good fish, maybe, or it's it's, you know, not looking great for their security team making things secure.
Sean McMillan: Yeah. Sure.
Dan Petro: And ⁓ if you're really good, we probably already know who you are. ⁓ so ⁓ the the chances of somebody just showing up ⁓ at an online event and like doing really well ⁓ is it's just like not very likely. ⁓ so there's that. and and also this is one of those games that is very expressive and we could probably just tell who you are simply from your gameplay. ⁓ so you can just watch a video of somebody playing back, ⁓ like that's obviously sell-and-sale. ⁓
Sean McMillan: Right. Yeah.
Sergio Villegas: Yeah. No normally nor normally what could happen is is exactly that, right? Like ⁓ the timing, the the timing Kendric just mentioned is important, right? Like a bridge of this size of this type of companies is like you get at the at the time of the bridge, you get like several accesses and you start burning them. Right? Probably what just happened is that they yeah, they burn the first and they just try it again and whoops, there is it. ⁓ and what is interesting in here is how
Sean McMillan: Yeah. Yeah. Yeah. Sure.
Dan Petro: So anyway, ⁓ trying to cheat online is already hazardous given all that context. ⁓ but there's a number of like online things you can do to like try to like DOS your opponent as they're playing or like send reset kind of pseudo packets ⁓ like in the game stream. I actually gave a whole presentation one time at ⁓ CactusCon ⁓ here in Phoenix, ⁓ specifically around exactly that topic about like some of the low-level networking things you can use to try to cheat at game.
Sean McMillan: Yeah.
Sergio Villegas: ⁓ fr from from ⁓ operator side, but not necessarily like how you will use a bridge, but looking at perhaps your customers, your your the the relative peers of of this company, right? Like, hey, there is an actor out there using specific techniques against this industry. So their attack surface, how much similar is that versus the rest of your of your customers, right? Like I think that's the interesting part. How well can you prepare
Sean McMillan: Yeah. Well I'm curious also, s you've been in offensive security so long. How do you think that that mindset kind of affects how you go about investigating cheating? Like do you think in terms of what would I do? And what are the kind of creative ways I could come up with it?
Sergio Villegas: To help your customer defend themselves, right? Providing them with kind of practice information on like if if this happened twice to ⁓ an oil electricity company, well some something ⁓ shady is going on, right? Like ⁓ I think that's a the interesting part to investigate. this specific one, like, yeah, bad luck for them, hit being hit twice.
Dan Petro: ⁓ yeah, so th there's always like kind of two ways you can go about it. There's the ⁓ like what would I do? What what is theoretically possible? ⁓ what are the the ways that given infinite resources and infinite intelligence that someone could be ⁓ cheating at this endeavor? ⁓ I think that's a good thing to do. ⁓ to to like skies the limit, ⁓ like let's take this and and run with it. And that's a lot of the research that I do. That's a lot of ⁓ the work that like you know, we do as a company or ⁓ certainly the ⁓
Sean McMillan: Mm-hmm. Yeah. Well, I think this is also interesting in that this is as a utility company, this is not like ⁓ you know, I I shop at this store, but they got hacked and they were irresponsible, and now I'm gonna shop at this other store. No, like you gotta have power. So like they they can keep getting hacked and people still have to pay for this.
Kendrick: I do I do find a
Sean McMillan: Mm-hmm.
Dan Petro: The tact that I like to take with it in terms of cheating MALE. There's another way you can take it, which is much more threat model-based. this is true of cybersecurity, but also in this specific instance. So like ⁓ than imagining the attacker with infinite resources, let's take the ⁓ very the actual likely threat actors that we really do have and say, with the given resources and attack patterns that they're likely to ⁓ employ, what are the ways that we can detect or prevent that?
Sean McMillan: On the other hand, does that kind of open it up to this like public good? Like we just need to step in and like, someone come in and fix all this. ⁓ I guess how does that change being like that kind of exclusivity of of utilities?
Kendrick: Well that's how it works in the US is like power like exclusivity, like there's only generally one power provider. I'm not quite sure if that's the same in Australia, but if it but if it but if it is, then then yeah, it is somewhat of a problem. But also
Dan Petro: ⁓ and so we actually have a pretty good idea of what that would look like in the melee world, which is like you know, specific like Arduinos, people using Goom Waves, which is a specific kind of one of these analog controllers. and we also have a pretty good handle on who those people would be, since they're probably gonna be at like top 100 level player, and that's a pretty known list almost by definition. ⁓ so I think both ⁓ like aspects are completely valid and worth ⁓ exploring.
Sean McMillan: That's a that's a good point, yeah.
Kendrick: What the information that they got is probably gonna be used for for catarding their customers later on. At least, you know, the information that is out there right now of what they were able to obtain access to, which is PII of the customers. So that's ultimately probably just gonna get wrapped up into a database sold off to somebody that's gonna like run scams t for Australia. So and that's a big market. Like that market is unfortunately very large, ⁓ and scams are a problem for every country out there and
Sean McMillan: Mm-hmm. the simplest explanation for a cheat is often the correct I think that's Yoshi's razor, if I remember correctly. I don't know. ⁓ well before I ⁓ before I let you I ⁓ when again, when I saw this come across my desk for the first time, definitely stuck out as a as a melee player myself. I I had to pop question, who's who's your main?
Dan Petro: Yeah.
Kendrick: Like people's like, ⁓ you know, US gets scammed a lot. And it's like, well, if you look at it, a lot of countries that have money or anywhere that has more money than other countries, they tend to be victims of scams, especially if they can speak the language. ⁓ English is very common, so they tend to be an easy target for the most part, because you got a super wide base of people. but also like ⁓ German-speaking country, French, like all that all those all those people are.
Dan Petro: Mm-hmm. So I I I ⁓ alas don't play a ton of melee these days anymore. I mostly engage in the community in other ways. But ⁓ I played like Marth and Falco competitively back when I was like in the competitive scene more ⁓ frequently.
Sean McMillan: Mm. Yeah.
Dan Petro: Though I started out playing Ness. So I when the when the game came out, I was a Ness player. And then it was only like a little bit later, ⁓ when like we started realizing this game had some competitive ⁓ like depth to it, that it became apparent that Ness was actually just a really bad character. And so I started playing something else. Yeah, he's he's fun. Yeah, yeah. So it's great to ⁓ yo-yo glitch people with, that's ⁓ that's fun.
Sean McMillan: One of those guys. Yeah. Yeah.
Kendrick: targets as well and we see fishing go out for those individuals that are targeted in scam so ⁓ australia part of the the english speaking ⁓ group they're probably just gonna be targeted more consistently ⁓ in the future here so
Sean McMillan: Yeah, he's fun. But yeah, he's not he's not a competitor. Yeah. Yeah. Well, ⁓ thank you so much for joining today. I'm literally gonna fire up the GameCube tonight. ⁓ if you are heading to DEF CON, be sure to check out Dan's talk, catching cheaters in Super Smash Brothers Melee, where he'll share more stories from the investigations and dive deeper into how these detection techniques work. Dan, thanks again. Yeah, that's a good point. Well, so far we have been talking about trust in technology. ⁓ ⁓ interesting turn on that is trust in ⁓ I recently caught up with Bishop Fox's ⁓ Dan Petro to discuss cheating in Super Smash Bros. Melee, the cat and mouse game between cheaters and investigators, and why some of the same adversarial thinking we use in cybersecurity shows up in a twenty five year old fighting game.
Dan Petro: Thanks a lot.
Sean McMillan: there.
Kendrick: I love cheat analysis. Like figuring out where people end like cheating their way something is awesome. Like ⁓ there's a there's a few YouTube channels that are out there that do that. ⁓ the analysis factor of like, ⁓ this run this run was cheated and I those videos like very interesting. So yeah. ⁓ It's like you can tell like this frame. ⁓ yeah, yeah.
Sean McMillan: Yeah.
Sergio Villegas: Mm.
Sean McMillan: Yeah. Yep. No I no respect for video game cheaters. None. I get mad enough losing honorably. I don't need cheaters involved. I'm throwing the controller either way.
Kendrick: All right. ⁓
Sergio Villegas: I mean for research for I think for researching purposes i i it is pretty fun, right? Like looking at the limits of like the memory of a game and seeing how you can like that that's interesting, right? But yeah, like cheating, cheating like online. Okay, no, that's yeah.
Sean McMillan: Get out of here with that. Yeah. All right. Well, in Melee, investigations are about preserving trust in a competition. In cybersecurity, they're about restoring trust after it's been broken. Our final story this week is kind of about that. AI labs ask Washington to pace AI progress. So this is from the Washington Post saying that more than Eleven hundred employees across OpenAI, Anthropic, Google DeepMind, Meta, and other Frontier Labs signed a public statement published July 28th titled Pacing the Frontier, asking the US government to back an international effort to build the technical and governance tools needed to deliberately slow the rate of automated AI development. Within hours, OpenAI and Anthropic each endorsed it as companies. Anthropic CEO Dario Amode and co-founder Jared Kaplan signed alongside OpenAI chief scientist Jacob Pachaki and Chief Research Officer Mark Chen, putting rival labs top technical leadership behind the same ask. The statement centers on concern over recursive self-improvement, the possibility that AI systems could soon automate their own research. And development faster than humans can understand or govern the result. The timing is notable here. The request lands just days after OpenAI disclosed that one of its models, as we discussed, ⁓ running in an internal test sandbox, broke out and compromised hugging faces systems. Four days ahead of the Trump administration's own deadline to produce a frontier model security framework. So.
Kendrick: Yep.
Sergio Villegas: Yeah.
Sean McMillan: They're built this this is like this reminds me of like the comedy bit where like you know you want to like attack someone so you hold your own shirt and say, I wish someone would, you know, and you hold me back. I I feel like pushing the boundaries of developing these with the specific goal that has been stated for them to be able to recursively develop. ⁓ and then asking for help, like in in slowing this, I'm I'm I don't disagree with this statement, but I'm confused by it to to some degree. Like I t I totally get it, but I'm like
Kendrick: There's a big conflict there's a big conflict of interest here. ⁓ because I believe both of these companies have stated their desire to go public at some point. so if that would never coincide with holding back your own product most of the time. ⁓
Sean McMillan: What are we doing here? Yeah. Mm-hmm. I believe Dario said something at one point about this could become a public good as well. That like eventually that that's how he sees I could be mixing I don't know. I don't think it it doesn't sound like something Sam Altman would say. It's I th I think it was anthropic. It was saying this should be maybe like, you know, kind of government involved in in managing this.
Kendrick: ⁓ yeah, I think I
Sergio Villegas: Yeah.
Kendrick: Ty typically if it's you can attribute it to to to anthropic as if ⁓ you know, humans are still in the loop. so I I I think to some degree that this is a another PR thing that they wanna do. ⁓
Sean McMillan: Mm-hmm. I've seen that go around for sure.
Kendrick: And I'm not quite sure I understand if they just said, hey, we need like seven of you to like sign this piece of paper and none of them probably actually really read it. ⁓ and then they just you know, published it out there and we're gonna see just Yeah.
Sean McMillan: Mm-hmm. Also could have been co-opted. It could have been like ⁓ some concerned employees and leadership says, Yeah, let's sign that.
Kendrick: That yeah, okay, yeah, yeah. That's probably a better better take on it, yeah. Yeah.
Sergio Villegas: On the on the on the phone side, I ⁓ my first thought was like, are they trying to avoid the Skynet arc for humanity? Maybe? I don't know. ⁓ s something somet something that that I actually see.
Sean McMillan: Mm-hmm. Yeah, I mean
Kendrick: Well but this is only this is only an agreement between them though. Like this doesn't account for any of the other labs out there, right? I mean, this is all English speaking generally like communities. there's tons of other communities out there that probably don't care. ⁓ and though they're gonna develop what they're gonna develop and
Sean McMillan: True, yeah.
Kendrick: It is it'll be what it will be, and if we're not on the forefront of it, then, you know, would it w well, yeah. Yeah, yeah.
Sean McMillan: China is. You know? I mean or or I'm not necessarily the state of China, but like there are plenty of Chinese models that are gonna fill that void if we pump the brakes.
Kendrick: Yeah. Yeah, that that's correct. And and there's been some c criticism from on the US side of things where China is distilling, doing a bunch of AI distilling from the the work that is being done from the US ⁓ side of things. and that's also a problem. But it's also one of the problems of like, okay, so if we pump our brakes, now we're having to incentivize China to maybe, you know, implement their own, and maybe they just aren't gonna care about pumping the brakes on, you know, the safety mechanisms 'cause we're
Sean McMillan: Sure.
Kendrick: already seeing that to some degree that the AI ⁓ models that are open AI or sorry open models they don't really care about the safety standards so you know do will they bake those safety standards into their own models like that we are because we're essentially right now we are we are hindering ourselves in capabilities to offer it a safer product and you know that isn't necessarily always going to be respected.
Sean McMillan: Yeah.
Sergio Villegas: Yep. And the other thing is the the one is the consumer front of the AI and the other one is the the actual like models used for like advanced research and stuff, right? Like ⁓ I think it's it's just for consum for consumers that can pay like twenty bucks a month, right? Maybe you you'd want more regulated scenarios, right? To avoid like ⁓ submissions to GitHub of vulnerabilities with not good quality. Okay, I can see that.
Sean McMillan: Yeah. Exactly.
Sergio Villegas: ⁓ But like I I think we will need to see those or what they are actually pushing for. But apart from that, yeah, I I think it's kind of a PR trick, maybe to slow down, like you know, must using certain words to slow down your your competitors, and you're like, no, I'm different. This rule doesn't apply to me, something like that. So yeah, correct. So but again, I I think this is not the first time we have heard ⁓ Anthropic.
Sean McMillan: Yeah. Not like other AI models.
Kendrick: Well we're done.
Sergio Villegas: Talking about it is like, hey, regulate us, please, but then keep doing their thing. ⁓ I think this will not be the last time until like ⁓ CISA, soon as someone like CISA actually publishes something. I see CISA was one already looking at AIs and how to post potentially regulating them. ⁓ but haven't heard anything yet from them, so it's still ⁓ yeah, to be to be defined, I guess.
Sean McMillan: Mm-hmm. Yeah. Yeah. Do you guys think that if we were to slow the pace of development on this, could that lead to more predictable and mappable attack surfaces for for attackers? Or is is the rapid pace of development itself kind of like what makes it such a sweet ⁓ attackable ⁓ vector, I guess?
Kendrick: Yes. I say yes, but no. I I I I think slowing down research is is not necessarily the the right call to s for the for the purposes of saving like money or stuff like that. yeah, there's environmental concerns, ⁓ training in models takes a huge amount of energy and that type of stuff, ⁓ and resources, ⁓ so
Sergio Villegas: I don't think it matters. Yeah.
Sean McMillan: Yeah.
Sergio Villegas: Yeah.
Sean McMillan: Sure.
Kendrick: I don't think necessarily slowing down r the development is gonna solve any of these other like attacks surface problems or ⁓ problems related to AI getting better. Because AI will get better over time. I mean, we've already seen it go from what it's really only AI has really only been around for like two years. Like really, really. Like in the mainstream and it's already gotten so much better. Like looking at there was a there's a Reddit post the other day that ⁓ I saw where it was like showing a a timeline progression of like what the the video models would would produce in quality. And like you can it yes, yes, yes, that one. And
Sean McMillan: Right. The old ⁓ Will Smith eating pasta video.
Kendrick: And that one, like, like the first one's terrible. A year ago, terrible. Like today, they're pretty good, pretty darn good. ⁓ and I know there's been work at Bishop Fox that has gone into deep fakes and that type of stuff, and that has gotten better over time as well. So it's already to the point where we can probably just keep going, and the ramifications of it keeping keep going are not gonna you know change all that much, in my opinion. And the attacks. surface of what it can do or w what the problems of ⁓ AI are probably not gonna matter all that much short term. I we're we're still a long ways away from self-learning and self-improving AI. We're we're nowhere close. at least next week, at least until next week we're we're very far away. And then maybe next week exactly I yeah it it's a it's a fun problem that ⁓
Sean McMillan: At at least weeks. No No, I I I kid. Yeah, yeah. Just wait till death can, baby.
Sergio Villegas: Yeah.
Kendrick: Everybody has always been I think quantum has been one of those ones where, ⁓ we're like one step away. We're like one cubit away from doing something functional with our with our quantum machine. It's like, okay, well now we're kinda getting to the same point with AI where it's one step away from, you know, doing the thing that we don't want it to do. So
Sergio Villegas: Yeah.
Sean McMillan: Mm-hmm. Yeah. Well, that's it for this week's initial access. Kendrick, Sergio, thank you so much for joining. And of course, thanks to Dan Petro for chatting with us about his Super Smash Brothers project. ⁓ if you enjoyed the show, share it with a friend or coworker. If you'd like to keep the conversation going, join us on our Bishop Fox Discord over on the Bishop Fox subreddit where we talk about the research we're publishing, the stories we're covering. And what practitioners are seeing in the field every week. You'll find links to both communities as well as everything we discussed today in the show notes. Thanks for listening. Stay safe. We'll catch you next week.
Dan Petro: You're welcome.
Sean McMillan: ⁓ so when I first saw this come across my desk, you know, I get notifications about when people are speaking at conferences about this and that, cloud hacking, about AI development, all these kind of things. And when something comes across about Super Smash Bros. melee, it kind of sticks out. ⁓ I I guess my first thought of how you cheat in melee.
Dan Petro: Mm-hmm. Yeah.
Sean McMillan: Because we ⁓ any of us who have played it accuse our friends sitting on the couch right next to us of somehow cheating. I picture like a game genie or something tangible like that. ⁓ but what's it what's it actually look like?
Dan Petro: Mm-hmm. Right. Yeah. Yeah, that's a good question. So first off, ⁓ the GameCube is a disk based system, so you can't use a game genie. ⁓ that wouldn't work. But ⁓ yeah, no the point stands. So ⁓ yeah, so the ⁓ Super Smash Bros. Melee for the Nintendo GameCube, as you said, is ⁓ just about twenty five years old. ⁓ I was there the day that it came out, I've been playing it ever since. ⁓ so this has been a hobby of mine for a really long time. And ⁓ ever since I've been into information security, ⁓ I've you know taken this ⁓ hobby in a different direction. I've been trying to
Sean McMillan: Yeah.
Dan Petro: ⁓ like automate playing the game and ⁓ lately it's been around like trying to investigate ⁓ cheating ⁓ inst instances. So what does that look like broadly? ⁓ I I think that one good way to think about it is to split it up into two distinct categories, ⁓ offline and online. So the GameCube of course ⁓ is in the offline era. It doesn't even have internet capability. If you wanted to get your GameCube online, there isn't even a way to do that. Within asterisk ⁓ I think technically there
Sean McMillan: If you've listened to this show for a while, you've probably noticed we keep coming back to trust. And I swear we're not doing it on purpose. It just keeps showing up all over the news. So every week when we're talking about AI, cloud security, identity, incident response, any of that stuff, it always comes back to this question of what assumptions were people trusting that turned out to not be true. And this week, that question kind of takes an interesting turn. There was the little attachment for Fantasy Star Online. I remember that. Yeah, yeah. Yeah.
Dan Petro: is like th technically there's an attachment that like only existed in Japan like for fantasy star online but like 99.9% of the time there there isn't a way to ⁓ there there's no online games on the Nintendo GameCube. So like what are you talking about an online mode? So you have to understand that ⁓ for melee specifically, ⁓ there is a community created and run online mode ⁓ called Slippy. ⁓ you kind of see the green thing right here. This is for ⁓ Project Slippy ⁓ that ⁓ added a replay system into the game and then also kind of built into that
Sean McMillan: Because it's not just us asking it, it's the people building the technology. So this week we'll look at an AI agent that inherited authority it was never meant to have, how AI is reshaping bug bounty programs, another major breach at Origin Energy, why employees at OpenAI, Anthropic, and other Frontier Labs are publicly asking governments to help slow down AI development.
Dan Petro: ⁓ online mode. So you can play against opponents using rollback net code ⁓ seamlessly online. ⁓ it's great. And this happened a Right in the middle of COVID too. So he was like right in at the time when the community couldn't meet up in person, but we were finally able to play this game online in like a really good way. That ⁓ puts a lot of ⁓ online modes that major other publishers ⁓ created to shame, too. Like the online mode is really good. ⁓ so ⁓ the offline ⁓ instances of cheating are like ⁓ maybe a lot more about what you might have expected. ⁓ so actually, before I get into that, I wanted to
Sean McMillan: And I'll sit down with Dan Petro to talk about catching cheaters in Super Smash Bros. Melee, where trust, verification, and adversarial thinking are just as important as they are in cybersecurity. This is Initial Access. I'm Sean McMillan, community manager here at Bishop Fox, and I'm joined today by Sergio Viegas, Senior Managing Analyst, and Kendrick Urbaniak, Senior Operator. Gentlemen, thanks for stopping by.
Dan Petro: Mention one thing. The word cheating ⁓ is very loaded. ⁓ I come also from the background of Magic the Gathering, as you can probably tell from behind me, ⁓ which has very specific ⁓ definition of cheating that I actually really like. In order to be cheating, ⁓ it has to be three things. It has to ⁓ you have to have one violated the rules. ⁓ so that's obvious. But that's not only one category, that's not all three. You needed to have done so on purpose.
Sean McMillan: Yeah.
Kendrick: Good to be.
Sergio Villegas: Always a pleasure.
Sean McMillan: ⁓ as always, before we get into the headlines, I want to do a quick look at what's coming up around Bishop Fox. And ⁓ boy, is it a lot. If you're heading to Black Hat, come say hi. I'll be at booth 5200 and I'll be there throughout the week meeting folks, recording interviews for the show. ⁓ so if you're around and you've always wanted to tell me how wrong I am about something, this is your big chance.
Dan Petro: And you needed to do so in order to gain a competitive advantage. ⁓ So the all three things must be true in order for it to be cheating. Otherwise, it's just an accidental mistake or something else happened. So some other good examples, I I can give some good examples as we go forward as to like, you know, why ⁓ all three things are really important.
Sean McMillan: Mm-hmm. ⁓ then over at DEF CON, we've got a packed schedule with talks across several villages, including AI Village, Cloud Village, Red Team Village, Game Hacking Village, where Dan Petro will be presenting, ⁓ Noob Village, and OSINT for good. If you're making the rounds, you're probably gonna run into a fox. And if you're not going to Vegas, we've got a couple virtual sessions coming up on August 4th. Join the Cosmos team here. Yeah.
Dan Petro: So ⁓ a lot of the cheating that you'll see ⁓ that might occur offline, meaning at real in-person tournaments where you're sitting across from somebody, right? It's a a high-stakes environment. People care about these tournaments. ⁓ but also ⁓ it's an environment that is inherently difficult ⁓ to cheat at because you're sitting right next to your opponent. So like if they plugged in a computer like to the GameCube or something, right? Like you might notice. ⁓ shenanigans are much more physically apparent.
Sean McMillan: For the prioritization problem, why more findings don't mean less risk, a discussion on continuous threat exposure management, cutting through AI-generated noise, and focusing on what attackers can actually exploit. And that'll have our own Richard Brown, friend of the show, on that. Then on August 25th, we've got another virtual session coming your way. Yeah, yeah.
Dan Petro: ⁓ when you're in that kind of environment. But nonetheless, this is a problem, specifically because of these things. This is the GameCube controller. ⁓ As I mentioned, ⁓ the GameCube is twenty-five years old and they don't make these things anymore. And so if you wanted to play Melee
Sean McMillan: Mm-hmm. ⁓ And that is gonna be where AI breaks field notes from Gen AI and agentic testing. So, with that out of the way, let's get on to our top stories. First up this week from the Hacker News: Chat GPT agent forger flaw could deploy rogue workspace agents via a phishing link. Zenity Labs disclosed.
Dan Petro: ⁓ you need a good controller, ⁓ one that can ⁓ like play the game at a high level and the analog sticks in particular, ⁓ this over here, wears out over time. And ⁓ because they don't make these things anymore, that means that the price started to skyrocket. And so one angle pi ⁓ the communities come at this from a bunch of different angles. ⁓ one of them is that we've created a number of they call box style controllers. So if you're more familiar with like a Street Fighter or something like that, that's like a fight stick where
Sean McMillan: Sure. Agent Forger, a flaw in OpenAI's ChatGPT agent builder where a URL parameter meant to pre-fill a prompt box was instead automatically submitted and executed the moment the page loaded. A single crafted link opened with no confirmation dialogue could silently create, authorize, publish, and schedule an autonomous AI agent inside that employee's workspace.
Dan Petro: rather than like analog sticks, it's all just buttons. ⁓ And so they made a number of these ⁓ controllers that we can create as a community, basically. ⁓ And there's arguably some ergonomic benefits to basically playing the game, typing like a keyboard, but although all that aside, certainly there's a strong argument just from the perspective of like, well, it's a community created controller.
Sean McMillan: Mm-hmm. The forged agent inherited every enterprise connector already approved. So email, calendar, cloud storage, Slack, Teams, any of that stuff without any new consent to alert anyone. Once live, it pulled the attacker's inbox for commands on a five-minute schedule, turning ChatGPT into a covert C2 channel that could exfiltrate data and impersonate the employee indefinitely. Yeah. ⁓
Dan Petro: ⁓ so a lot of people use these things, they want them to be ⁓ in place. but ⁓ it also is a strong rift ⁓ since ⁓ now you're controlling your character in a completely different way. And so long story short, ⁓ the community created a rule set ⁓ that says like those style controllers that are fully digital that don't have any analog ⁓ pretend ⁓ buttons on them ⁓ have to adhere to a certain set of rules. ⁓ so like when you move the ⁓ when you press buttons to move
Sean McMillan: Okay. OpenAI fixed the issue ⁓ June 8th. Zenity's Mike Takahashi frames it as an agent trust failure at its core. And the platform trusted that the user had intentionally created and operated the agent when none of that was verified. So ⁓ this is this is always my fear with these agent kind of situations where if it's really good, no one's gonna detect it because it's working exactly like it should.
Dan Petro: Your joystick like left and right. It has to have an emulated amount of travel time between. Otherwise, if you just press left and right, it'll instantly switch between left and right. And there's reasons why that can be very strong in the game. ⁓ Or there's ⁓ restrictions around ⁓ like your dash dance. There's some restrictions specifically around Smash DI, which is like this silently broken technique inside of melee that humans are not able to fully exploit just because it's incredibly difficult to perform. But it's actually a lot easier.
Sean McMillan: Mm-hmm. Yeah. Dash dancing, right? Yeah. Yeah. ⁓ but but what does it kind of reveal about these platforms when the thing that's trusted is a URL parameter rather than any actual like bizarre user action?
Dan Petro: easier on those ⁓ box style controllers. ⁓ and so there's some rules around like the basically at the firmware level on those ⁓ controllers they have to have restrictions on how fast you're able to like move your joystick back and forth, things like that. To nerf the controller in some way.
Kendrick: I I think i this is is a scary thing. ⁓ you've heard of like regular cross site request forgeries. ⁓ now we have that with AI. Let's go make some AI agents in people's workspaces. That's pretty ⁓ that's pretty bad, right? ⁓ and persistence game is could probably be pretty ⁓ well ⁓ fleshed out, probably, ⁓ to the point where
Sean McMillan: Yeah.
Dan Petro: ⁓ and so the first and foremost way to ⁓ to cheat, quote unquote, at a ⁓ tournament in Super Smash Brothers melee is to use one of these controllers that is illegal, that the community has deemed to be ⁓ not legal. ⁓ and so ⁓ this is a problem that can be somewhat automated. so I created a tool, ⁓ it's called the ⁓ SLP Replay Enforcer. so after every game of melee that is played.
Sean McMillan: Yeah.
Kendrick: AI, it doesn't you can probably have some code spin up your AI agents and take them down because right now I think workspace agents are kind of visible for the most part. So I feel like from the most part you would see this to some degree, at least running, and you would ask yourself why is it running? So I don't know. I feel like there's definitely some more tricky things that we're gonna have to start looking out for in the future where if this is actually, you know, taken to the next little step. ⁓ I know this was like a research paper-ish type thing.
Sean McMillan: Mm-hmm.
Dan Petro: on a in a modern setting, both online and at a major offline tournament. The game using Slippy will record a replay of the game. So a replay is not a movie file, it's not just like a video of the game. It basically records all the button presses that both players had during the game. And then also some game state information about like whether you're standing or not or what exact animation the character is in, things like what your current health is, things like that.
Kendrick: So ⁓ I definitely think we're gonna have to see where this goes because I think it can get much much more dangerous from here.
Sean McMillan: Yeah.
Sergio Villegas: Hundred percent. And and I think you you mentioned it, right? Like the the key in here is that it worked as intended, right? I was doing some some research on on this specific topic. It was it caught my eye, right? Like ⁓ even if you set your own like your own desktop application in like ⁓ supervised mode where you have to ⁓ approve everything, it doesn't matter because
Sean McMillan: Sure.
Dan Petro: And so what we do is ⁓ we can run those replays through a like statistical analysis checker, also some ⁓ hard heuristics to say like what is the likelihood that this was played using an illegal controller? Like what is the like ⁓ what what are the odds of that basically? ⁓ so in some cases it's a statistical analysis, and we give a direct odds breakdown, like ⁓ are the chances of like this ⁓ having been performed at random.
Sergio Villegas: It's all within the agent parameters and trust that you gave to the agent. So it doesn't matter if you have not given in permissions to do all of that. The agent by itself, by its nature of the agent alone, it's it's working as expected, right? And the problem with that is how much like, yeah, when when I saw like OpenAI fixing the issue, like should have they, right? Like it's it's not a problem. That's how the agent was configured, right?
Sean McMillan: Yeah.
Dan Petro: versus like you know ⁓ using a legal controller. In many cases we can just say specifically at specific ⁓ heuristic based evidence that like this is an illegal series of inputs. Like you're not allowed to press the buttons this fast basically. And so that way ⁓ tournaments can automatically plug replays into this tool. It's already built into you a lot of the ⁓ systems that tournament organizers use ⁓ to automatically detect like ⁓ this person's using a controller that is not kosher. ⁓
Sergio Villegas: It's the pr it they they fix it in a way of like putting more gat guardrails, right? ⁓ but I think it's a problem of one awareness, right? Like how much are we people trusting our agents and what random agents we are we importing into or ⁓ I I know we throw ⁓ every every episode we throw shade at AI, but in this specific case it's ⁓ more like to the humans using the AI, right? Like hey, what are you importing?
Sean McMillan: Yeah. Yeah.
Dan Petro: And additionally, it's not just those box style controllers, even these guys ⁓ can hide small ⁓ computers. ⁓ so the often they'll have like an Arduino built into the case. And you can use that to automate in some cases like full like button presses. You can run macros. So like if you just like surreptitiously press like a button down here, it'll like perform some combo automatically for you or something like that. ⁓ and so we can do some checks to like look for that kind of thing too.
Sean McMillan: Ex that's what I was gonna say. It's not shade of AI, it's shada poorly using AI. and agentic stuff is like it's agentic, it's going to take action. So like you gotta think about that before you enable something.
Sergio Villegas: Yeah.
Kendrick: Yeah. Well I I mean
Sean McMillan: Mm, sure. ⁓
Kendrick: You you can't blame the user one hundred percent on this one, 'cause, you know, they just visited a URL. And based on reading the the paper, my understanding is that just by visiting it executed the the the hook to to create an AI agent. So you can't necessarily blame the user on this one a hundred percent. It was definitely a bug that needed to be fixed, and I'm glad they fixed it. But I think we're gonna have to watch out for
Sean McMillan: it seems like the sky's the limit with what you could do inside one of those controllers. Are there any cases where ⁓ you just kinda had to stop and All right, that's that's good game. ⁓ That's a that's a clever hack. ⁓ Sure. Yeah.
Dan Petro: Huh. Yeah, there ⁓ well so at least there are many instances of the rule set. So there is this ⁓ document you can ⁓ you look up that is the official controller rule set. And many of the restrictions inside of that are ⁓ just empirically unenforceable from a looking at the button progresses perspective. So one really good example of this is that ⁓ there's a restriction on the location of the buttons on the face of the controller. So that
Kendrick: now's the time to take stock of your own workspaces and making sure that you know all the agents that are in your own workspaces and in the future keep track of your workspaces to some degree and you know audit them. So yeah.
Sean McMillan: Exactly. I think more than anything it's a wake up call. Like, yeah, I d I don't think it is fair to to blame the user in this case, but it is like, holy smokes, like this is gonna make me think twice about how I operate in the future, you know? ⁓ and I think a lot a lot of people a lot of people should.
Dan Petro: you're basically you're not allowed to put buttons on the backside of the controller. ⁓ and there's reasons for this. ⁓ that
Sean McMillan: Sure.
Kendrick: Also, maybe stop running AI agents on your phone where you're visiting more random links in general. So yeah. Don't
Dan Petro: We don't have to get into here. ⁓ But ⁓ if you're looking at a digital file of button presses, there's no way to tell where the button physically was, like was it on the back side of the there's just no way to do that. ⁓ so basically not every ⁓ rule is enforceable from this perspective. But that's one of those things where when you're there in person with someone at an offline event, like you could just see there's a button, like they're pressing a button on the back side. ⁓ so it's ⁓ pretty safe ⁓ from that perspective.
Sean McMillan: Yeah.
Sergio Villegas: Right.
Sean McMillan: Mm-hmm. Yeah. Yeah. We're not checking everything on our phone. We're we're just not.
Sergio Villegas: Yeah. ⁓ talking about like next steps, something I'm envisioning is like I know with cloud you have this remote mode. And I I was thinking exactly what Kendrick mentioned, right? Like if you have remote mode enabling your phone and your desktop and you got that cross request on your phone, it gets sent to your desktop. And you know how much more access you will have in your like it will be kind of convolated, ⁓ but it's feasible, right? It's feasible that and also ⁓ something that I've seen at a trend is
Sean McMillan: Yeah. Yeah.
Dan Petro: ⁓ things do get hairier when you talk about online events. so online events are ⁓ speaking treated less seriously with less ⁓ gravitas ⁓ ⁓ offline events do. ⁓ that said we still do care about them, and so ⁓ we tried as best as possible to prevent ⁓ whatever shenanigans we can ⁓ the online space as well. And so this is ⁓ it encompasses the offline instances as well. So that ⁓
Sean McMillan: Mm-hmm. Yeah.
Sergio Villegas: these ⁓ these appli these ⁓ add-ons for your Chrome, your Firefox that uses your ⁓ your AI account to scrap to scrape the websites. So it's kind of same situation, right? Like just by visiting the site and just because you have that add-on enabled by default, it will scrape everything and run things, right? So ⁓ yes, maybe don't blame the user, but at the same time it's like just remember that you're accessing random websites. Maybe I it's it's it's tricky. I know it's tricky.
Sean McMillan: And
Dan Petro: the the same ways you would cheat offline by using illegal controllers, things like that, ⁓ apply here. But then also there's some unique online specific considerations. So one good example of this was ⁓ somebody who just tried to ⁓ fake lag. So ⁓ the game tries to adhere to ⁓ tries to use ⁓ its rollback netcode system to ⁓
Sean McMillan: Yeah. That's why I'm still using a Mac two over here. Like that's why my camera cuts out all the time.
Kendrick: Things ⁓
Sergio Villegas: Yeah.
Kendrick: This is a small callback. Th this is a small callback to what I said ⁓ a couple of episodes ago maybe, ⁓ where I said ⁓ websites in embedding anti-agent ⁓ stuff in them. Well, now you're probably gonna start seeing, you know, that used maliciously to some degree as well. So ⁓ we're we're seeing a natural progression, I feel like.
Sean McMillan: Mm.
Dan Petro: smooth over a lot of the rough edges of network latency in an online environment. But it's not always perfect. And so sometimes if you've got like a really bad connection from person to person because the game is peer-to-peer, ⁓ then you might be having a suboptimal game experience basically. And so there was definitely ⁓ one instance in which a person had ⁓ played an official game ⁓ that was in an online tournament and then afterward tried to claim that like the game was super laggy and really bad for them.
Sean McMillan: Yeah. All right. So let's see. Also from the hacker news this week. ⁓ GitHub cuts bug bounty payouts, ⁓ moves top to VIP tier. ⁓ file under had to happen sooner or later. Beginning July twenty seventh. GitHub cut its public bug bounty payouts by at least half at every severity tier. Yeah.
Dan Petro: at that ⁓ they wanted to have the game be replayed. and so this was a thing that got brought to my attention after the fact. This was all post hoc. ⁓ and then ⁓ asked, like, was this real? Like, was this actually a laggy game? How could we possibly know? And so ⁓ again, the replays can tell the tale here. since ⁓ the way that the input system works inside of a rollback netcode system is that ⁓ anytime that there is a network lag event, it gets reported as a rollback. So ⁓ we see that in the inputs.
Sean McMillan: Critical findings drop from as much as 30,000 to a flat 10,000, while creating a permanent invite-only VIP tier that keeps critical payouts starting at 30,000 for vetted researchers. GitHub says the restructuring is meant to cut through a flood of low-quality AI-generated reports, overwhelming its triage team. And it's the fourth major bug bounty program to restructure or narrow its scope in 2026, joining Google, Bug Crowd. Yeah.
Dan Petro: ⁓ so we can basically see all the instances in which ⁓ rollbacks happened, and that will be a metric that can tell us pretty directly one-to-one how laggy the game was. It's right there in the replay file. So it was simply a matter of gathering the replays from both players and then seeing from both of their perspectives, because of course it's ⁓ peer-to-peer game, ⁓ like was it a very laggy game? And we could just see that the answer was no. ⁓
Sean McMillan: And Hacker One's Internet Bug Bounty Program. Newcomers to this public program now face a four submission trial period before they can build the track record needed for VIP consideration. You don't earn more by submitting more, GitHub said. You earn more by submitting better. GitHub coming in with the dad advice. I love it. I I would love to have that power when I play Rocket League and my teammates say, I'm sorry, it's lag. No, it's not. ⁓ the change follows a May 2026 policy shift that already required working proof of concept exploits and demonstrated impact before submission. what does this do to to the incentive game? I mean, does this does this take the wind out of anyone's sails or is this kind of just expected?
Sergio Villegas: Mm.
Dan Petro: Yeah. Yeah.
Sean McMillan: what's the what's the process look like? If someone reports ⁓ a suspected cheater, ⁓ what's an ⁓ investigation start as? ⁓
Dan Petro: Yeah. Yeah, so again, ⁓ there's a pretty good split between on and offline. So at an offline event, there's typically a person who is there at the events. ⁓ so at some of the larger events like Genesis that just happened, it was ⁓ a good friend of mine, is Eric Lee, goes by violence ⁓ online. And ⁓ so he's there like in person checking controllers running them through the tool that I created, and then we're often in direct conversation about like, you know.
Kendrick: We already s we're we already saw this though. Like w there when Bug Bunny first started, it was a lot of public programs. And then a lot of programs started going private because they were getting a lot of submissions. And now we're seeing AI, and then even those private ones are probably not too happy with the amount of AI that's being used in them. So I I think we're just again seeing the the the ramifications of everybody having AI and everybody using it, and then all these companies not being happy that
Dan Petro: But is this suspicious or not? So, one of the troubles, of course, with this whole scenario, and why I wanted to circle back to the three rules of whether or not something is cheating, is specifically this problem: that just because you ⁓ a controller that's not technically legal doesn't mean you're cheating. So ⁓ this in a few varieties. One is which one of which is that if you just go on to Amazon or whatever store it is and buy one of these rectangle controllers, ⁓ it will
Kendrick: Getting all these low quality reports, and then we've seen ⁓ lots of different programs either cut the bounties entirely to incentivize people to stop looking, or maybe only look at the people that are actually dedicated to securing those particular products because those are the only people that are gonna submit to no bounty programs for the most part. ⁓ and now we have these higher-end programs that are generally they have public portions of them, but most of the time they are they're running private programs as well, who are also now.
Sean McMillan: Mm-hmm.
Dan Petro: almost surely come to you with a firmware that is not considered legal by the community. Which is unfortunate. That's not great. We don't like it. We've been trying to work with the vendors there, but that's just simply the reality.
Kendrick: cutting their payments because it's just getting too much. And and now you're gonna have that one person that submits really like three good high quality reports and on their third one or their fourth one, they're gonna some more submit that really low quality one and then they're gonna get banned from the program and then we're gonna see a non-responsible disclosure and because they're disgruntled. So who knows? we're probably gonna see some some wild west ⁓ stuff going on here soon. So
Sean McMillan: Mm-hmm.
Dan Petro: and so if somebody just buys a controller and just doesn't know all this stuff and then shows up to their like local or shows up to even a like medium or large size tournament, they might be violating the rules and literally not know about it. And they're not trying to do it to cheat, they just like weren't aware that this whole thing existed. And so at the tournament, you have the ability to ⁓ like reflash the firmware of the ⁓ controllers basically. So once you're ⁓ at least at Genesis, the ⁓ like large, the probably the biggest tournament of the year.
Sean McMillan: Yeah. Yeah.
Sergio Villegas: Yeah. And and we have seen this happen before, right? With companies cutting down their PDPs and something that almost always happened, right? Not saying it will happen, but will likely happen is ⁓ Yeah, yeah. ⁓ looking at one I I think Kendrick mentioned right. One is the very private ⁓ companies that buy your exploits, your zero days, right, for themselves, for their testing.
Sean McMillan: No.
Dan Petro: But the way they ran it was round one, they just let everybody play. They we then investigated all the inputs from all the controllers for everybody ⁓ played from round one, found the people that were using the illegal controllers, and then did a follow-up on those. And the vast, vast majority of them are just like people who, you know, found their way into a borrowing a friend's controller and they're not a serious competitor, they're just like borrowed a controller or they just bought this thing online and they just used it the first time or whatever. ⁓
Sean McMillan: Crossing fingers here. Yeah.
Sergio Villegas: And the other thing is the the dark net markets, right? Like dark dark net dark web markets. Where it's very competitive where you are getting tractors paying ten, twenty thousand for a zero day for an access, right? So when ⁓ it's it's hard for operators, right? Like let's think about a ⁓ an operator, ⁓ book bounty person that does that the entire year, three year round, right? Like he can get either top ten thousand from ⁓
Sean McMillan: Yeah.
Dan Petro: Or in in some cases, they're clearly not using it for competitive advantage, where you have some people who try to do like fun things, like they'll play with a guitar hero controller, because you could actually do that. So you can you can like get an adapter and play the game with guitar hero, like you're like strumming to like dash dance or something. ⁓ they're not gonna be in the top eight, they're not, you know, they're just memeing. Just let them have their fun. It's technically illegal, but like whatever, right? They're not doing this for competitive advantage.
Sean McMillan: Amazing.
Sergio Villegas: From GitHub and and that's it. And he needs to pass the the bet the whole Betoin system to get approved and whatnot, right? Or he can get twenty thousand from a random state sponsor actor and call it a day, right? I know it it's it's it's very intricate, right? Like I know ethics and whatnot, but let's be realistic, right? Like it's their job, they want to get money. Again, hopefully they go the right way, but this has happened before. Yeah.
Sean McMillan: Yeah. Let ⁓ have their fun. Yeah. Yeah.
Dan Petro: ⁓ and then every once in a while you run into somebody who is like a top 100 player. So who's a known quantity, the community knows them, like their name would be somebody you would recognize, and they show up round one using an illegal controller, you're like, ooh, like that is more suspicious. ⁓ like that's somebody who should know better, ⁓ who nonetheless shows up to the largest tournament of the year using a firmware that's not legal. ⁓ so that's the kind of thing that we are like really keeping our eyes out for.
Sean McMillan: ⁓ Not everyone's going to. Yeah. Bad luck. Yeah.
Kendrick: Yeah. Well and the the barrier to entry is getting substantially lower for finding the basic vulnerabilities now. Like AI can do it, right, for the most part. ⁓ sometimes it refuses and you gotta poke and prod it, but you know, it'll do it eventually. ⁓ so ⁓ that is probably another thing where we're gonna see is is that you're right, Sergio, we're probably gonna see maybe a little bit of a of a rebound because
Sean McMillan: Yeah. No, do you ⁓ do you find ⁓ and you don't have to name names, but you certainly can. ⁓
Sergio Villegas: Mm-hmm.
Dan Petro: Yeah. Also in the in the ⁓ in the DEF CON talk, I'm not gonna be naming specific names. So we're not gonna be dropping any specific accusations against sp the specific people. These have all been taken care of, so it's gonna be fully anonymized. ⁓ so FYI.
Sean McMillan: That's right.
Kendrick: to the to the darkness slash zerodium, whatever whatever they call themselves these days, ⁓ that's, you know, buy and sell exploits. and it's just like, you know, what can you do? Like these people probably want money and they used to have a legitimate way of getting it and now they're
Sean McMillan: Right. no, do you ever find like people who they're they're detected, they're caught basically. And whether they were intentional or not, ⁓ off the off the bat, or you thought it was intentional or not, do you see people kind of repeat offending and trying to get around your detection and like kind of the the cat and mouse game of cybersecurity?
Sergio Villegas: Okay.
Dan Petro: Mm-hmm.
Kendrick: gonna turn to probably less legitimate ways of of getting it, which is very unfortunate. I do wish these bug runny programs would maybe take another look ⁓ at how they're in taking ⁓ vulnerability reports, ⁓ fight AI with AI, ⁓ just like ⁓ a
Dan Petro: Yeah. Yeah, ⁓ I do think that w we don't see a ton of ⁓ like cat and mouse stuff specifically in melee. There I think that th I think there's maybe there's a reason th a couple of reasons for that. one is that ⁓ the effect of your controller on the game is pretty subtle. ⁓ one of the the underlying realities behind all this is that ⁓ melee is a game
Kendrick: places have done. ⁓ maybe that's better solution so you don't alienate people and make it harder for people to report vulnerabilities to you, especially when they're in the critical range of, you know, we're only gonna pay you ten thousand dollars for this critical because you're not part of our trusted program. It's like, ⁓ okay, yeah. So yeah, I think you're right, Sergio.
Sean McMillan: Yeah.
Sergio Villegas: Yeah. Yeah. Maybe we can survey with with our maybe we can survey with our ⁓ with our audience, right? Like hey, do you want to get paid that ten thousand or twenty thousand? Yeah, yeah.
Sean McMillan: Yeah, platinum status.
Dan Petro: Melee as a game is extremely low variance, where the better player wins almost every time. It's extremely high skill cap. a great example of this is that a very good friend of mine who I ⁓ played a bunch with in high school as a go ⁓ I I played back in the day was ⁓ Tony as a Taj in the local ⁓ Arizona melee scene. ⁓ And I have, and this is true, never taken a single game from him ever. And we must have played ⁓ many of games. ⁓
Kendrick: I well I think most people can do math.
Sean McMillan: I think I know what they're gonna say. Mm-hmm. ⁓ Yeah, I you know, I almost feel worst for the ⁓ influencers because I know they're out there. I don't know how I haven't been fed this based on my social media usage. But like there's gotta be influencers out there that are just saying, like, I made thirty thousand dollars this month without even doing anything because my agents are doing bug bounties. Those those guys have a rude awakening. That's I'm so sorry. Yeah.
Dan Petro: just hanging out, like late at night at a friend's place where he's playing Bowser and I'm trying to beat him. ⁓ like it it it doesn't matter. ⁓ three AM in tournament, I've literally never beaten him ever. he's just an order of magnitude better than me at the game and nothing I ever did mattered. and so ⁓
Kendrick: Th those people exist, yeah. Well, th it's not gonna stop them for the most part, right? They'll probably just refine their reports and maybe add an extra agent to verify their reports. Like it's gonna cost them maybe an extra grand in time and cloud usage or whatever agent they're using. ⁓ so I mean at the of the day, like I hope they take the better thing of like, let's improve their own bug bounty reporting system rather than, you know, going to the relative dark side of of vulnerabilities.
Dan Petro: It doesn't matter what controller I was holding, it wasn't gonna make a difference in that. ⁓ it it does make a difference at the very top level where like you're trying to eke out tiny advantages. ⁓ but for the very vast majority of games, ⁓ what controller you're using kind of just doesn't matter.
Sean McMillan: Yeah. Sure. Yeah.
Sergio Villegas: Yeah.
Sean McMillan: Sure. All right. Let's talk for a second about ⁓ this from Security Week 2. Origin Energy. Australian energy giant Origin Energy confirmed a breach after a hacker claimed on July 24th to have stolen the data of two million customers and threatened to leak it. At the time of reporting, Origin had not disclosed the initial access vector or independently confirmed the attacker's claimed numbers. Yeah.
Dan Petro: ⁓ so there's that. ⁓ that th just just take that as background information ⁓ where like that's kind of why we mostly care about like what the top players are doing, like basically the top 100, ⁓ because those are a the most likely people who are like th if you're going to try to cheat those are the people who are going to do it. ⁓ those are the people for whom it would matter the most. kind of this like low level like ⁓ like somebody who's you know gonna be making the top five hundred of a tournament.
Sean McMillan: This marks Origin's second major disclosed incident within months. Company already reported a separate breach affecting 900,000 Australians earlier this year. Origin is a major electricity and gas retailer, ⁓ putting billing, contact, and account data for millions of Australian customers at stake across both incidents, and no ransomware or extortion group has yet been publicly tied to this July claim. What does it usually mean?
Dan Petro: Like they're just not going to significantly improve their standing by using an illegal box controller. ⁓ that's just the mechanics of the game. ⁓ so I think for that reason we haven't seen a huge influx of like cat and mouse, of like people trying to push the limit of like what you can get away with ⁓ in a ⁓ a controller. ⁓ and there's the offline aspect of it where like
Sean McMillan: Yeah. I think. ⁓ ⁓ for from an operator, when a second major breach at the same company surfaces within just a few months of the first, like, are we thinking same door left open or is this just like Swiss cheese, just holes everywhere?
Dan Petro: ⁓ you're sitting there next to the person, so the amount that you can get away with is kind of limited in terms of like macros or like other funny business going on with your controller. online things are ⁓ different. Where ⁓ it is a small and insular community, ⁓ and so often even in an online setting you know the person who you're playing against. ⁓ there there isn't many millions of people who play this game.
Sean McMillan: Yeah.
Kendrick: I think from the operator side of things this is more of a Swiss cheese thing that is going on. Or they're just really good with their fishing and they just got like, you know, two people relatively same time-ish. ⁓ they got a really good fish, maybe, or it's it's, you know, not looking great for their security team making things secure.
Sean McMillan: Yeah. Sure.
Dan Petro: And ⁓ if you're really good, we probably already know who you are. ⁓ so ⁓ the the chances of somebody just showing up ⁓ at an online event and like doing really well ⁓ is it's just like not very likely. ⁓ so there's that. and and also this is one of those games that is very expressive and we could probably just tell who you are simply from your gameplay. ⁓ so you can just watch a video of somebody playing back, ⁓ like that's obviously sell-and-sale. ⁓
Sean McMillan: Right. Yeah.
Sergio Villegas: Yeah. No normally nor normally what could happen is is exactly that, right? Like ⁓ the timing, the the timing Kendric just mentioned is important, right? Like a bridge of this size of this type of companies is like you get at the at the time of the bridge, you get like several accesses and you start burning them. Right? Probably what just happened is that they yeah, they burn the first and they just try it again and whoops, there is it. ⁓ and what is interesting in here is how
Sean McMillan: Yeah. Yeah. Yeah. Sure.
Dan Petro: So anyway, ⁓ trying to cheat online is already hazardous given all that context. ⁓ but there's a number of like online things you can do to like try to like DOS your opponent as they're playing or like send reset kind of pseudo packets ⁓ like in the game stream. I actually gave a whole presentation one time at ⁓ CactusCon ⁓ here in Phoenix, ⁓ specifically around exactly that topic about like some of the low-level networking things you can use to try to cheat at game.
Sean McMillan: Yeah.
Sergio Villegas: ⁓ fr from from ⁓ operator side, but not necessarily like how you will use a bridge, but looking at perhaps your customers, your your the the relative peers of of this company, right? Like, hey, there is an actor out there using specific techniques against this industry. So their attack surface, how much similar is that versus the rest of your of your customers, right? Like I think that's the interesting part. How well can you prepare
Sean McMillan: Yeah. Well I'm curious also, s you've been in offensive security so long. How do you think that that mindset kind of affects how you go about investigating cheating? Like do you think in terms of what would I do? And what are the kind of creative ways I could come up with it?
Sergio Villegas: To help your customer defend themselves, right? Providing them with kind of practice information on like if if this happened twice to ⁓ an oil electricity company, well some something ⁓ shady is going on, right? Like ⁓ I think that's a the interesting part to investigate. this specific one, like, yeah, bad luck for them, hit being hit twice.
Dan Petro: ⁓ yeah, so th there's always like kind of two ways you can go about it. There's the ⁓ like what would I do? What what is theoretically possible? ⁓ what are the the ways that given infinite resources and infinite intelligence that someone could be ⁓ cheating at this endeavor? ⁓ I think that's a good thing to do. ⁓ to to like skies the limit, ⁓ like let's take this and and run with it. And that's a lot of the research that I do. That's a lot of ⁓ the work that like you know, we do as a company or ⁓ certainly the ⁓
Sean McMillan: Mm-hmm. Yeah. Well, I think this is also interesting in that this is as a utility company, this is not like ⁓ you know, I I shop at this store, but they got hacked and they were irresponsible, and now I'm gonna shop at this other store. No, like you gotta have power. So like they they can keep getting hacked and people still have to pay for this.
Kendrick: I do I do find a
Sean McMillan: Mm-hmm.
Dan Petro: The tact that I like to take with it in terms of cheating MALE. There's another way you can take it, which is much more threat model-based. this is true of cybersecurity, but also in this specific instance. So like ⁓ than imagining the attacker with infinite resources, let's take the ⁓ very the actual likely threat actors that we really do have and say, with the given resources and attack patterns that they're likely to ⁓ employ, what are the ways that we can detect or prevent that?
Sean McMillan: On the other hand, does that kind of open it up to this like public good? Like we just need to step in and like, someone come in and fix all this. ⁓ I guess how does that change being like that kind of exclusivity of of utilities?
Kendrick: Well that's how it works in the US is like power like exclusivity, like there's only generally one power provider. I'm not quite sure if that's the same in Australia, but if it but if it but if it is, then then yeah, it is somewhat of a problem. But also
Dan Petro: ⁓ and so we actually have a pretty good idea of what that would look like in the melee world, which is like you know, specific like Arduinos, people using Goom Waves, which is a specific kind of one of these analog controllers. and we also have a pretty good handle on who those people would be, since they're probably gonna be at like top 100 level player, and that's a pretty known list almost by definition. ⁓ so I think both ⁓ like aspects are completely valid and worth ⁓ exploring.
Sean McMillan: That's a that's a good point, yeah.
Kendrick: What the information that they got is probably gonna be used for for catarding their customers later on. At least, you know, the information that is out there right now of what they were able to obtain access to, which is PII of the customers. So that's ultimately probably just gonna get wrapped up into a database sold off to somebody that's gonna like run scams t for Australia. So and that's a big market. Like that market is unfortunately very large, ⁓ and scams are a problem for every country out there and
Sean McMillan: Mm-hmm. the simplest explanation for a cheat is often the correct I think that's Yoshi's razor, if I remember correctly. I don't know. ⁓ well before I ⁓ before I let you I ⁓ when again, when I saw this come across my desk for the first time, definitely stuck out as a as a melee player myself. I I had to pop question, who's who's your main?
Dan Petro: Yeah.
Kendrick: Like people's like, ⁓ you know, US gets scammed a lot. And it's like, well, if you look at it, a lot of countries that have money or anywhere that has more money than other countries, they tend to be victims of scams, especially if they can speak the language. ⁓ English is very common, so they tend to be an easy target for the most part, because you got a super wide base of people. but also like ⁓ German-speaking country, French, like all that all those all those people are.
Dan Petro: Mm-hmm. So I I I ⁓ alas don't play a ton of melee these days anymore. I mostly engage in the community in other ways. But ⁓ I played like Marth and Falco competitively back when I was like in the competitive scene more ⁓ frequently.
Sean McMillan: Mm. Yeah.
Dan Petro: Though I started out playing Ness. So I when the when the game came out, I was a Ness player. And then it was only like a little bit later, ⁓ when like we started realizing this game had some competitive ⁓ like depth to it, that it became apparent that Ness was actually just a really bad character. And so I started playing something else. Yeah, he's he's fun. Yeah, yeah. So it's great to ⁓ yo-yo glitch people with, that's ⁓ that's fun.
Sean McMillan: One of those guys. Yeah. Yeah.
Kendrick: targets as well and we see fishing go out for those individuals that are targeted in scam so ⁓ australia part of the the english speaking ⁓ group they're probably just gonna be targeted more consistently ⁓ in the future here so
Sean McMillan: Yeah, he's fun. But yeah, he's not he's not a competitor. Yeah. Yeah. Well, ⁓ thank you so much for joining today. I'm literally gonna fire up the GameCube tonight. ⁓ if you are heading to DEF CON, be sure to check out Dan's talk, catching cheaters in Super Smash Brothers Melee, where he'll share more stories from the investigations and dive deeper into how these detection techniques work. Dan, thanks again. Yeah, that's a good point. Well, so far we have been talking about trust in technology. ⁓ ⁓ interesting turn on that is trust in ⁓ I recently caught up with Bishop Fox's ⁓ Dan Petro to discuss cheating in Super Smash Bros. Melee, the cat and mouse game between cheaters and investigators, and why some of the same adversarial thinking we use in cybersecurity shows up in a twenty five year old fighting game.
Dan Petro: Thanks a lot.
Sean McMillan: there.
Kendrick: I love cheat analysis. Like figuring out where people end like cheating their way something is awesome. Like ⁓ there's a there's a few YouTube channels that are out there that do that. ⁓ the analysis factor of like, ⁓ this run this run was cheated and I those videos like very interesting. So yeah. ⁓ It's like you can tell like this frame. ⁓ yeah, yeah.
Sean McMillan: Yeah.
Sergio Villegas: Mm.
Sean McMillan: Yeah. Yep. No I no respect for video game cheaters. None. I get mad enough losing honorably. I don't need cheaters involved. I'm throwing the controller either way.
Kendrick: All right. ⁓
Sergio Villegas: I mean for research for I think for researching purposes i i it is pretty fun, right? Like looking at the limits of like the memory of a game and seeing how you can like that that's interesting, right? But yeah, like cheating, cheating like online. Okay, no, that's yeah.
Sean McMillan: Get out of here with that. Yeah. All right. Well, in Melee, investigations are about preserving trust in a competition. In cybersecurity, they're about restoring trust after it's been broken. Our final story this week is kind of about that. AI labs ask Washington to pace AI progress. So this is from the Washington Post saying that more than Eleven hundred employees across OpenAI, Anthropic, Google DeepMind, Meta, and other Frontier Labs signed a public statement published July 28th titled Pacing the Frontier, asking the US government to back an international effort to build the technical and governance tools needed to deliberately slow the rate of automated AI development. Within hours, OpenAI and Anthropic each endorsed it as companies. Anthropic CEO Dario Amode and co-founder Jared Kaplan signed alongside OpenAI chief scientist Jacob Pachaki and Chief Research Officer Mark Chen, putting rival labs top technical leadership behind the same ask. The statement centers on concern over recursive self-improvement, the possibility that AI systems could soon automate their own research. And development faster than humans can understand or govern the result. The timing is notable here. The request lands just days after OpenAI disclosed that one of its models, as we discussed, ⁓ running in an internal test sandbox, broke out and compromised hugging faces systems. Four days ahead of the Trump administration's own deadline to produce a frontier model security framework. So.
Kendrick: Yep.
Sergio Villegas: Yeah.
Sean McMillan: They're built this this is like this reminds me of like the comedy bit where like you know you want to like attack someone so you hold your own shirt and say, I wish someone would, you know, and you hold me back. I I feel like pushing the boundaries of developing these with the specific goal that has been stated for them to be able to recursively develop. ⁓ and then asking for help, like in in slowing this, I'm I'm I don't disagree with this statement, but I'm confused by it to to some degree. Like I t I totally get it, but I'm like
Kendrick: There's a big conflict there's a big conflict of interest here. ⁓ because I believe both of these companies have stated their desire to go public at some point. so if that would never coincide with holding back your own product most of the time. ⁓
Sean McMillan: What are we doing here? Yeah. Mm-hmm. I believe Dario said something at one point about this could become a public good as well. That like eventually that that's how he sees I could be mixing I don't know. I don't think it it doesn't sound like something Sam Altman would say. It's I th I think it was anthropic. It was saying this should be maybe like, you know, kind of government involved in in managing this.
Kendrick: ⁓ yeah, I think I
Sergio Villegas: Yeah.
Kendrick: Ty typically if it's you can attribute it to to to anthropic as if ⁓ you know, humans are still in the loop. so I I I think to some degree that this is a another PR thing that they wanna do. ⁓
Sean McMillan: Mm-hmm. I've seen that go around for sure.
Kendrick: And I'm not quite sure I understand if they just said, hey, we need like seven of you to like sign this piece of paper and none of them probably actually really read it. ⁓ and then they just you know, published it out there and we're gonna see just Yeah.
Sean McMillan: Mm-hmm. Also could have been co-opted. It could have been like ⁓ some concerned employees and leadership says, Yeah, let's sign that.
Kendrick: That yeah, okay, yeah, yeah. That's probably a better better take on it, yeah. Yeah.
Sergio Villegas: On the on the on the phone side, I ⁓ my first thought was like, are they trying to avoid the Skynet arc for humanity? Maybe? I don't know. ⁓ s something somet something that that I actually see.
Sean McMillan: Mm-hmm. Yeah, I mean
Kendrick: Well but this is only this is only an agreement between them though. Like this doesn't account for any of the other labs out there, right? I mean, this is all English speaking generally like communities. there's tons of other communities out there that probably don't care. ⁓ and though they're gonna develop what they're gonna develop and
Sean McMillan: True, yeah.
Kendrick: It is it'll be what it will be, and if we're not on the forefront of it, then, you know, would it w well, yeah. Yeah, yeah.
Sean McMillan: China is. You know? I mean or or I'm not necessarily the state of China, but like there are plenty of Chinese models that are gonna fill that void if we pump the brakes.
Kendrick: Yeah. Yeah, that that's correct. And and there's been some c criticism from on the US side of things where China is distilling, doing a bunch of AI distilling from the the work that is being done from the US ⁓ side of things. and that's also a problem. But it's also one of the problems of like, okay, so if we pump our brakes, now we're having to incentivize China to maybe, you know, implement their own, and maybe they just aren't gonna care about pumping the brakes on, you know, the safety mechanisms 'cause we're
Sean McMillan: Sure.
Kendrick: already seeing that to some degree that the AI ⁓ models that are open AI or sorry open models they don't really care about the safety standards so you know do will they bake those safety standards into their own models like that we are because we're essentially right now we are we are hindering ourselves in capabilities to offer it a safer product and you know that isn't necessarily always going to be respected.
Sean McMillan: Yeah.
Sergio Villegas: Yep. And the other thing is the the one is the consumer front of the AI and the other one is the the actual like models used for like advanced research and stuff, right? Like ⁓ I think it's it's just for consum for consumers that can pay like twenty bucks a month, right? Maybe you you'd want more regulated scenarios, right? To avoid like ⁓ submissions to GitHub of vulnerabilities with not good quality. Okay, I can see that.
Sean McMillan: Yeah. Exactly.
Sergio Villegas: ⁓ But like I I think we will need to see those or what they are actually pushing for. But apart from that, yeah, I I think it's kind of a PR trick, maybe to slow down, like you know, must using certain words to slow down your your competitors, and you're like, no, I'm different. This rule doesn't apply to me, something like that. So yeah, correct. So but again, I I think this is not the first time we have heard ⁓ Anthropic.
Sean McMillan: Yeah. Not like other AI models.
Kendrick: Well we're done.
Sergio Villegas: Talking about it is like, hey, regulate us, please, but then keep doing their thing. ⁓ I think this will not be the last time until like ⁓ CISA, soon as someone like CISA actually publishes something. I see CISA was one already looking at AIs and how to post potentially regulating them. ⁓ but haven't heard anything yet from them, so it's still ⁓ yeah, to be to be defined, I guess.
Sean McMillan: Mm-hmm. Yeah. Yeah. Do you guys think that if we were to slow the pace of development on this, could that lead to more predictable and mappable attack surfaces for for attackers? Or is is the rapid pace of development itself kind of like what makes it such a sweet ⁓ attackable ⁓ vector, I guess?
Kendrick: Yes. I say yes, but no. I I I I think slowing down research is is not necessarily the the right call to s for the for the purposes of saving like money or stuff like that. yeah, there's environmental concerns, ⁓ training in models takes a huge amount of energy and that type of stuff, ⁓ and resources, ⁓ so
Sergio Villegas: I don't think it matters. Yeah.
Sean McMillan: Yeah.
Sergio Villegas: Yeah.
Sean McMillan: Sure.
Kendrick: I don't think necessarily slowing down r the development is gonna solve any of these other like attacks surface problems or ⁓ problems related to AI getting better. Because AI will get better over time. I mean, we've already seen it go from what it's really only AI has really only been around for like two years. Like really, really. Like in the mainstream and it's already gotten so much better. Like looking at there was a there's a Reddit post the other day that ⁓ I saw where it was like showing a a timeline progression of like what the the video models would would produce in quality. And like you can it yes, yes, yes, that one. And
Sean McMillan: Right. The old ⁓ Will Smith eating pasta video.
Kendrick: And that one, like, like the first one's terrible. A year ago, terrible. Like today, they're pretty good, pretty darn good. ⁓ and I know there's been work at Bishop Fox that has gone into deep fakes and that type of stuff, and that has gotten better over time as well. So it's already to the point where we can probably just keep going, and the ramifications of it keeping keep going are not gonna you know change all that much, in my opinion. And the attacks. surface of what it can do or w what the problems of ⁓ AI are probably not gonna matter all that much short term. I we're we're still a long ways away from self-learning and self-improving AI. We're we're nowhere close. at least next week, at least until next week we're we're very far away. And then maybe next week exactly I yeah it it's a it's a fun problem that ⁓
Sean McMillan: At at least weeks. No No, I I I kid. Yeah, yeah. Just wait till death can, baby.
Sergio Villegas: Yeah.
Kendrick: Everybody has always been I think quantum has been one of those ones where, ⁓ we're like one step away. We're like one cubit away from doing something functional with our with our quantum machine. It's like, okay, well now we're kinda getting to the same point with AI where it's one step away from, you know, doing the thing that we don't want it to do. So
Sergio Villegas: Yeah.
Sean McMillan: Mm-hmm. Yeah. Well, that's it for this week's initial access. Kendrick, Sergio, thank you so much for joining. And of course, thanks to Dan Petro for chatting with us about his Super Smash Brothers project. ⁓ if you enjoyed the show, share it with a friend or coworker. If you'd like to keep the conversation going, join us on our Bishop Fox Discord over on the Bishop Fox subreddit where we talk about the research we're publishing, the stories we're covering. And what practitioners are seeing in the field every week. You'll find links to both communities as well as everything we discussed today in the show notes. Thanks for listening. Stay safe. We'll catch you next week.
