WEBVTT
1
00:00:00.000 --> 00:00:04.128
[Max] On Friday, OpenAI shipped what it calls the strongest AI model ever built.
2
00:00:04.128 --> 00:00:08.892
And then, in the same breath, it told almost everyone they can't use it yet.
3
00:00:08.892 --> 00:00:14.609
Not because of a bug. Not because of the price. Because the government has to clear you first.
4
00:00:14.609 --> 00:00:19.690
The best AI on earth just grew a velvet rope. And the bouncer works for Washington.
5
00:00:36.609 --> 00:00:43.119
[Max] Welcome to Context Window. Your weekly window into everything worth knowing in AI security.
6
00:00:43.119 --> 00:00:48.233
Quick thing before we start. Full transparency. This podcast is AI-generated.
7
00:00:48.233 --> 00:00:54.278
Two AI voices, a script written by another AI agent, the whole thing.
8
00:00:54.278 --> 00:01:01.717
But every story, every take, every bit of here-is-what-this-actually-means is curated and verified by Assaf Nakash.
9
00:01:01.717 --> 00:01:07.761
He builds AI security products at one of the world's largest security platforms.
10
00:01:07.761 --> 00:01:15.665
The robots do the talking, but a human who lives this stuff makes sure we're not hallucinating.
11
00:01:15.915 --> 00:01:24.838
[Nina] And if the robots ever start vetting which humans are allowed to hear this episode, you'll know we've gone full frontier model.
12
00:01:24.838 --> 00:01:28.082
Until then, anything sounds off, let Assaf know.
13
00:01:28.332 --> 00:01:30.604
[Max] Three things this week. The big one.
14
00:01:30.604 --> 00:01:36.120
OpenAI built its most powerful model yet and then handed the guest list to the US government.
15
00:01:36.120 --> 00:01:38.067
The counter-move that's already showing up.
16
00:01:38.067 --> 00:01:42.934
Free models you can download are catching the paid frontier faster than they ever have.
17
00:01:42.934 --> 00:01:52.992
And in the speed round, where the money is flowing, and a copy-paste mistake that put the same hole in the AI servers behind some of the biggest names in tech.
18
00:01:53.242 --> 00:01:57.100
[Nina] That last one is my favorite kind of bug. Nobody got clever.
19
00:01:57.100 --> 00:02:01.602
Somebody just hit copy, then paste, a few hundred times. Let's get into it.
20
00:02:05.001 --> 00:02:11.986
[Max] Okay, top story. Friday, June twenty-sixth. OpenAI announces a new family of frontier models. Three of them.
21
00:02:11.986 --> 00:02:18.971
Sol, the one it calls its strongest yet. Terra for everyday work. And Luna, the cheaper option.
22
00:02:18.971 --> 00:02:23.079
Normal product launch so far. And then comes the asterisk.
23
00:02:23.079 --> 00:02:30.475
At the US government's request, the first release goes only to a small group of vetted, trusted partners.
24
00:02:30.475 --> 00:02:34.584
Everyone else gets access later. In the coming weeks, supposedly.
25
00:02:34.834 --> 00:02:43.472
[Nina] Hold on. Vetted by whom, exactly, and vetted for what? What does a trusted partner even mean here?
26
00:02:43.722 --> 00:02:45.226
[Max] The wording matters here.
27
00:02:45.226 --> 00:02:53.128
It means OpenAI previewed the model to the government before launch and let it help decide who's in the first wave.
28
00:02:53.128 --> 00:02:58.771
Think of an early-access list, except the people approving the list don't work at OpenAI.
29
00:02:58.771 --> 00:03:02.157
They work for the government. That's the new part.
30
00:03:02.407 --> 00:03:04.784
[Nina] Okay, but be precise with me.
31
00:03:04.784 --> 00:03:11.914
Two weeks ago a rival, Anthropic, had two of its newest models pulled offline by a government order.
32
00:03:11.914 --> 00:03:15.876
That was a hard ban. Is this the same thing?
33
00:03:16.126 --> 00:03:22.508
[Max] No, and that's the whole story. That's the distinction I need everyone to hold onto.
34
00:03:22.508 --> 00:03:29.741
Anthropic got an order. An export-control directive that forced a shutdown, no choice involved. This is voluntary.
35
00:03:29.741 --> 00:03:43.783
Earlier this month the White House signed an executive order that asks AI labs to preview their most powerful models with the government and let it weigh in on who gets early access.
36
00:03:43.783 --> 00:03:50.165
Nobody forced OpenAI. It walked its flagship over to Washington and accepted a guest list.
37
00:03:50.415 --> 00:03:58.775
[Nina] See, my instinct is that voluntary is the less scary version. They opted in. They can opt out.
38
00:03:59.025 --> 00:04:04.766
[Max] And I think it's the opposite, and here's why. A one-time emergency order is an incident.
39
00:04:04.766 --> 00:04:07.637
You can argue it was a special case.
40
00:04:07.637 --> 00:04:13.737
But a lab choosing, on its own, to submit its best model for government sign-off before launch?
41
00:04:13.737 --> 00:04:18.761
That's not an incident. That's a norm forming. The hard order made the headlines.
42
00:04:18.761 --> 00:04:22.709
The soft norm is the thing that actually changes the industry.
43
00:04:22.959 --> 00:04:29.274
[Nina] That's a fair reframe. The ban is the exception. The volunteering is the precedent.
44
00:04:29.524 --> 00:04:36.264
[Max] Exactly. And what makes it land harder. OpenAI doesn't even like the arrangement.
45
00:04:36.264 --> 00:04:40.412
It publicly pushed back on its own restriction.
46
00:04:40.412 --> 00:04:48.707
It called this a short-term step and said government pre-approval should not become the long-term default.
47
00:04:48.707 --> 00:04:58.038
It literally warned that gating keeps the best tools away from defenders, developers, and enterprises who need them.
48
00:04:58.288 --> 00:05:01.115
[Nina] Which is a genuinely strange thing to watch.
49
00:05:01.115 --> 00:05:06.767
The company that built the gate is standing next to it arguing the gate shouldn't exist.
50
00:05:06.767 --> 00:05:08.180
And building it anyway.
51
00:05:08.430 --> 00:05:12.400
[Max] Right. And for the security crowd specifically, one detail.
52
00:05:12.400 --> 00:05:22.988
OpenAI says Sol is its most capable cybersecurity model ever, and that it's better at helping defenders fix vulnerabilities than at carrying out attacks.
53
00:05:22.988 --> 00:05:27.400
That's OpenAI's own claim, so take it as a claim.
54
00:05:27.400 --> 00:05:33.136
But it tells you why a government suddenly cares who gets the keys.
55
00:05:33.386 --> 00:05:34.554
[Nina] Put it together.
56
00:05:34.554 --> 00:05:42.737
Frontier AI is sliding from a product you just buy into a capability somebody signs off on before it reaches you.
57
00:05:42.737 --> 00:05:49.362
That's a lot closer to how we regulate advanced chips, or weapons, than how software normally ships.
58
00:05:49.612 --> 00:05:56.066
[Max] That's the takeaway. And if you build anything on top of these models, the question just got real.
59
00:05:56.066 --> 00:06:02.161
Which model is under my product, and could it be gated tomorrow? That used to be hypothetical.
60
00:06:02.161 --> 00:06:05.029
This week it became an architecture review item.
61
00:06:08.428 --> 00:06:10.843
[Max] Alright, let's shift gears.
62
00:06:11.093 --> 00:06:16.501
[Nina] I'm taking the first one, because it's the other half of the same story.
63
00:06:16.501 --> 00:06:22.680
Wall off the frontier, and you don't freeze the field. You hand the challengers a market.
64
00:06:22.680 --> 00:06:34.267
The free, downloadable models, the ones you don't rent, the ones you just grab and run on your own hardware, are catching the paid frontier faster than they ever have.
65
00:06:34.517 --> 00:06:36.468
[Max] How close are we actually talking?
66
00:06:36.718 --> 00:06:44.815
[Nina] On a respected independent ranking, the leading open model now sits about six points behind the top paid ones.
67
00:06:44.815 --> 00:06:52.060
That's the smallest that gap has ever been. A year ago it was roughly twice as wide.
68
00:06:52.060 --> 00:06:56.322
Now, important caveat. Six points behind is not caught up.
69
00:06:56.322 --> 00:07:03.141
On the hardest problems, the paid models are still clearly ahead. This is narrowing, not matching.
70
00:07:03.391 --> 00:07:06.178
[Max] And the part everyone in Washington is going to notice.
71
00:07:06.427 --> 00:07:11.749
[Nina] Several of the strongest open models come from outside the US. Chinese labs, specifically.
72
00:07:11.749 --> 00:07:15.549
So the capability the gate is trying to fence in?
73
00:07:15.549 --> 00:07:20.870
A lot of it is already over the fence, and it didn't ask permission.
74
00:07:21.120 --> 00:07:25.579
[Max] Which is a beautiful segue to where the money went this week.
75
00:07:25.579 --> 00:07:30.780
Two raises, and they're both bets on this exact shift. Start with the why.
76
00:07:30.780 --> 00:07:38.954
When countries decide they can't depend on a model somebody else can switch off, they start wanting to own the whole thing.
77
00:07:39.204 --> 00:07:41.014
[Nina] Own it. Not rent it.
78
00:07:41.264 --> 00:07:41.771
[Max] Right.
79
00:07:41.771 --> 00:07:53.916
A startup called Dream raised two hundred sixty million dollars, at a three billion dollar valuation, to sell governments what it calls sovereign AI.
80
00:07:53.916 --> 00:08:01.002
Infrastructure a nation fully owns and controls, instead of renting from a foreign provider.
81
00:08:01.252 --> 00:08:07.645
[Nina] A three billion dollar valuation for AI you never have to ask anyone's permission to run.
82
00:08:07.645 --> 00:08:10.442
That's the market pricing the fear directly.
83
00:08:10.442 --> 00:08:20.431
And the founder is the former head of NSO Group, the spyware company, which is its own kind of signal about who sees this coming.
84
00:08:20.680 --> 00:08:24.074
[Max] Second raise, completely different gap, same readiness problem underneath.
85
00:08:24.074 --> 00:08:30.860
A startup called NewCore came out of stealth with sixty-six million dollars to give AI agents real identities.
86
00:08:30.860 --> 00:08:32.368
Here's why that matters.
87
00:08:32.368 --> 00:08:38.401
The tools that decide who's allowed to do what inside a company were built for humans.
88
00:08:38.401 --> 00:08:41.040
Logins, roles, permissions, all designed around people.
89
00:08:41.040 --> 00:08:49.335
Then everyone deployed armies of autonomous agents that act on their own, and those tools just don't have a slot for them.
90
00:08:49.585 --> 00:08:52.066
[Nina] And that's not a theoretical gap.
91
00:08:52.066 --> 00:09:04.890
In a survey of a hundred and sixty-plus security leaders, run by an agent-security firm, most are rolling out agents, but only about a quarter have real controls built for them.
92
00:09:04.890 --> 00:09:10.268
One in five has already had a security incident caused by an agent.
93
00:09:10.268 --> 00:09:16.472
So NewCore is selling a fix for a hole companies have already started falling into.
94
00:09:16.722 --> 00:09:22.043
[Max] Last one in the round, and it's a good one. Completely different flavor.
95
00:09:22.043 --> 00:09:26.954
This week, Microsoft ran its BlueHat IL conference over in Tel Aviv.
96
00:09:26.954 --> 00:09:31.456
One of the biggest security-research gatherings anywhere outside the United States.
97
00:09:31.456 --> 00:09:36.366
Hundreds of researchers, and this year the program was wall-to-wall AI security.
98
00:09:36.616 --> 00:09:40.448
[Nina] And the headline talk revisited something with a great name. ShadowMQ.
99
00:09:40.448 --> 00:09:44.975
Walk me through it, because this is the copy-paste bug from the open.
100
00:09:45.225 --> 00:09:52.608
[Max] Picture this. One team writes an insecure way of passing data between programs that serve an AI model.
101
00:09:52.608 --> 00:09:58.349
A real flaw, the kind that lets an attacker take the machine over remotely.
102
00:09:58.349 --> 00:10:05.321
And then, instead of one team having that bug, everyone copies the code. Almost word for word.
103
00:10:05.321 --> 00:10:13.934
So the exact same remote-takeover hole ends up in the AI-serving software at Meta, at NVIDIA, even at Microsoft, among others.
104
00:10:13.934 --> 00:10:16.804
One mistake, pasted across the whole ecosystem.
105
00:10:17.054 --> 00:10:20.670
[Nina] Now, honesty check. That research is not brand new.
106
00:10:20.670 --> 00:10:25.892
It was first disclosed late last year, by an outside firm, Oligo Security.
107
00:10:25.892 --> 00:10:33.124
What's this week is the conference dissecting it for the broader community. We're not calling it breaking news.
108
00:10:33.373 --> 00:10:37.702
[Max] Right, and the lesson is what makes it worth the airtime.
109
00:10:37.702 --> 00:10:44.787
The soft underbelly of AI isn't always the model. Sometimes it's the boring plumbing that serves the model.
110
00:10:44.787 --> 00:10:49.116
And nobody audited the plumbing because everybody assumed somebody else did.
111
00:10:51.716 --> 00:10:54.301
[Max] Alright. Curator's Corner. The if-you-listen-to-nothing-else moment.
112
00:10:54.301 --> 00:10:59.904
And Assaf's title for it this week is the title of the episode.
113
00:10:59.904 --> 00:11:02.490
The level playing field is ending.
114
00:11:02.740 --> 00:11:08.840
[Max] Here's the argument. For a few genuinely remarkable years, the best AI on earth was a commodity.
115
00:11:08.840 --> 00:11:17.453
A student in Nairobi, a startup in Tel Aviv, and a Fortune 500 in New York could all rent the exact same frontier model.
116
00:11:17.453 --> 00:11:23.912
Same terms, by the token. That flatness was easy to take for granted. It was also historically weird.
117
00:11:23.912 --> 00:11:26.424
And this week, it started to end.
118
00:11:26.673 --> 00:11:38.238
[Nina] Because when the strongest model ships first to a government-approved guest list, and a rival's models can be switched off worldwide by one directive, AI stops behaving like software.
119
00:11:38.488 --> 00:11:41.054
[Max] And starts behaving like a strategic asset.
120
00:11:41.054 --> 00:11:45.453
Governed the way nations govern weapons, and the chips that build them.
121
00:11:45.453 --> 00:11:52.419
The way he sees it, if the pattern holds, the strongest countries get privileged access to the strongest models.
122
00:11:52.669 --> 00:11:54.851
[Nina] Here comes the equal and opposite.
123
00:11:55.101 --> 00:11:58.862
[Max] Exactly. Every force you put on a system pushes back.
124
00:11:58.862 --> 00:12:04.127
Wall off the frontier, and you don't freeze the field. You fund its challengers.
125
00:12:04.127 --> 00:12:10.897
We just watched it happen. Open models six points off the frontier. Sovereign-AI money raised the same week.
126
00:12:10.897 --> 00:12:14.281
The gate doesn't stop the water. It reroutes it.
127
00:12:14.532 --> 00:12:20.399
[Nina] But let me push on the framing, because I think it's softer than the episode title.
128
00:12:20.399 --> 00:12:25.899
OpenAI says this is voluntary, and temporary, and it's publicly against making it the norm.
129
00:12:25.899 --> 00:12:32.132
So is the era really over? Or is this a direction, and we're calling it a destination?
130
00:12:32.381 --> 00:12:38.976
[Max] That's fair. It's a trajectory, not a verdict. Nothing here is locked in. I'll give you that.
131
00:12:39.226 --> 00:12:47.168
[Nina] Good. Because the part that does land for me isn't the geopolitics. It's the security sting underneath it.
132
00:12:47.418 --> 00:12:48.206
[Max] Say it.
133
00:12:48.456 --> 00:12:55.879
[Nina] The capability everyone's busy fencing off is already loose. An autonomous AI agent recently ran a complete intrusion.
134
00:12:55.879 --> 00:13:02.063
Break-in, all the way to a stolen database, in under an hour. No human driving.
135
00:13:02.313 --> 00:13:05.408
[Max] Now look at the board you're actually playing on.
136
00:13:05.408 --> 00:13:08.845
The frontier model is gated away from the compliant defender.
137
00:13:08.845 --> 00:13:14.346
The one with logins, invoices, and an audit trail, who'd use it to protect a company.
138
00:13:14.346 --> 00:13:23.628
It is not gated away from the attacker who just downloads a free model that's nearly caught up, and runs it on a machine nobody can reach.
139
00:13:23.628 --> 00:13:28.784
And barely a quarter of the companies running agents are ready for any of it.
140
00:13:29.034 --> 00:13:35.768
[Nina] Which means the gate goes up exactly where the good guys live, and the bad guys walk around it.
141
00:13:36.018 --> 00:13:39.114
[Max] That's the whole thing. So forget the disclaimers.
142
00:13:39.114 --> 00:13:43.758
Here's where the bug actually lives, and what you do about it.
143
00:13:43.758 --> 00:13:51.110
Write down which model is under each product you ship, and what breaks if it gets gated next month.
144
00:13:51.110 --> 00:13:57.302
And stop assuming restricted means unavailable to attackers. Because the level playing field wasn't just fair.
145
00:13:57.302 --> 00:14:02.721
It was a defense. And we're dismantling it exactly where we'll miss it most.
146
00:14:06.120 --> 00:14:09.306
[Max] That's Context Window for the week of June twenty-ninth.
147
00:14:09.306 --> 00:14:19.216
The frontier grew a guest list, the free models kept gaining, and somewhere a copy-paste bug is still serving models at three of the biggest names in tech.
148
00:14:19.466 --> 00:14:21.850
[Nina] The full edition is on LinkedIn.
149
00:14:21.850 --> 00:14:30.989
Every source, every number, and the action items we didn't have time for, read by security leaders who do this for a living.
150
00:14:30.989 --> 00:14:33.770
That's where the real show notes live.
151
00:14:34.019 --> 00:14:35.274
[Max] See you next week.
1
00:00:00.000 --> 00:00:04.128
[Max] On Friday, OpenAI shipped what it calls the strongest AI model ever built.
2
00:00:04.128 --> 00:00:08.892
And then, in the same breath, it told almost everyone they can't use it yet.
3
00:00:08.892 --> 00:00:14.609
Not because of a bug. Not because of the price. Because the government has to clear you first.
4
00:00:14.609 --> 00:00:19.690
The best AI on earth just grew a velvet rope. And the bouncer works for Washington.
5
00:00:36.609 --> 00:00:43.119
[Max] Welcome to Context Window. Your weekly window into everything worth knowing in AI security.
6
00:00:43.119 --> 00:00:48.233
Quick thing before we start. Full transparency. This podcast is AI-generated.
7
00:00:48.233 --> 00:00:54.278
Two AI voices, a script written by another AI agent, the whole thing.
8
00:00:54.278 --> 00:01:01.717
But every story, every take, every bit of here-is-what-this-actually-means is curated and verified by Assaf Nakash.
9
00:01:01.717 --> 00:01:07.761
He builds AI security products at one of the world's largest security platforms.
10
00:01:07.761 --> 00:01:15.665
The robots do the talking, but a human who lives this stuff makes sure we're not hallucinating.
11
00:01:15.915 --> 00:01:24.838
[Nina] And if the robots ever start vetting which humans are allowed to hear this episode, you'll know we've gone full frontier model.
12
00:01:24.838 --> 00:01:28.082
Until then, anything sounds off, let Assaf know.
13
00:01:28.332 --> 00:01:30.604
[Max] Three things this week. The big one.
14
00:01:30.604 --> 00:01:36.120
OpenAI built its most powerful model yet and then handed the guest list to the US government.
15
00:01:36.120 --> 00:01:38.067
The counter-move that's already showing up.
16
00:01:38.067 --> 00:01:42.934
Free models you can download are catching the paid frontier faster than they ever have.
17
00:01:42.934 --> 00:01:52.992
And in the speed round, where the money is flowing, and a copy-paste mistake that put the same hole in the AI servers behind some of the biggest names in tech.
18
00:01:53.242 --> 00:01:57.100
[Nina] That last one is my favorite kind of bug. Nobody got clever.
19
00:01:57.100 --> 00:02:01.602
Somebody just hit copy, then paste, a few hundred times. Let's get into it.
20
00:02:05.001 --> 00:02:11.986
[Max] Okay, top story. Friday, June twenty-sixth. OpenAI announces a new family of frontier models. Three of them.
21
00:02:11.986 --> 00:02:18.971
Sol, the one it calls its strongest yet. Terra for everyday work. And Luna, the cheaper option.
22
00:02:18.971 --> 00:02:23.079
Normal product launch so far. And then comes the asterisk.
23
00:02:23.079 --> 00:02:30.475
At the US government's request, the first release goes only to a small group of vetted, trusted partners.
24
00:02:30.475 --> 00:02:34.584
Everyone else gets access later. In the coming weeks, supposedly.
25
00:02:34.834 --> 00:02:43.472
[Nina] Hold on. Vetted by whom, exactly, and vetted for what? What does a trusted partner even mean here?
26
00:02:43.722 --> 00:02:45.226
[Max] The wording matters here.
27
00:02:45.226 --> 00:02:53.128
It means OpenAI previewed the model to the government before launch and let it help decide who's in the first wave.
28
00:02:53.128 --> 00:02:58.771
Think of an early-access list, except the people approving the list don't work at OpenAI.
29
00:02:58.771 --> 00:03:02.157
They work for the government. That's the new part.
30
00:03:02.407 --> 00:03:04.784
[Nina] Okay, but be precise with me.
31
00:03:04.784 --> 00:03:11.914
Two weeks ago a rival, Anthropic, had two of its newest models pulled offline by a government order.
32
00:03:11.914 --> 00:03:15.876
That was a hard ban. Is this the same thing?
33
00:03:16.126 --> 00:03:22.508
[Max] No, and that's the whole story. That's the distinction I need everyone to hold onto.
34
00:03:22.508 --> 00:03:29.741
Anthropic got an order. An export-control directive that forced a shutdown, no choice involved. This is voluntary.
35
00:03:29.741 --> 00:03:43.783
Earlier this month the White House signed an executive order that asks AI labs to preview their most powerful models with the government and let it weigh in on who gets early access.
36
00:03:43.783 --> 00:03:50.165
Nobody forced OpenAI. It walked its flagship over to Washington and accepted a guest list.
37
00:03:50.415 --> 00:03:58.775
[Nina] See, my instinct is that voluntary is the less scary version. They opted in. They can opt out.
38
00:03:59.025 --> 00:04:04.766
[Max] And I think it's the opposite, and here's why. A one-time emergency order is an incident.
39
00:04:04.766 --> 00:04:07.637
You can argue it was a special case.
40
00:04:07.637 --> 00:04:13.737
But a lab choosing, on its own, to submit its best model for government sign-off before launch?
41
00:04:13.737 --> 00:04:18.761
That's not an incident. That's a norm forming. The hard order made the headlines.
42
00:04:18.761 --> 00:04:22.709
The soft norm is the thing that actually changes the industry.
43
00:04:22.959 --> 00:04:29.274
[Nina] That's a fair reframe. The ban is the exception. The volunteering is the precedent.
44
00:04:29.524 --> 00:04:36.264
[Max] Exactly. And what makes it land harder. OpenAI doesn't even like the arrangement.
45
00:04:36.264 --> 00:04:40.412
It publicly pushed back on its own restriction.
46
00:04:40.412 --> 00:04:48.707
It called this a short-term step and said government pre-approval should not become the long-term default.
47
00:04:48.707 --> 00:04:58.038
It literally warned that gating keeps the best tools away from defenders, developers, and enterprises who need them.
48
00:04:58.288 --> 00:05:01.115
[Nina] Which is a genuinely strange thing to watch.
49
00:05:01.115 --> 00:05:06.767
The company that built the gate is standing next to it arguing the gate shouldn't exist.
50
00:05:06.767 --> 00:05:08.180
And building it anyway.
51
00:05:08.430 --> 00:05:12.400
[Max] Right. And for the security crowd specifically, one detail.
52
00:05:12.400 --> 00:05:22.988
OpenAI says Sol is its most capable cybersecurity model ever, and that it's better at helping defenders fix vulnerabilities than at carrying out attacks.
53
00:05:22.988 --> 00:05:27.400
That's OpenAI's own claim, so take it as a claim.
54
00:05:27.400 --> 00:05:33.136
But it tells you why a government suddenly cares who gets the keys.
55
00:05:33.386 --> 00:05:34.554
[Nina] Put it together.
56
00:05:34.554 --> 00:05:42.737
Frontier AI is sliding from a product you just buy into a capability somebody signs off on before it reaches you.
57
00:05:42.737 --> 00:05:49.362
That's a lot closer to how we regulate advanced chips, or weapons, than how software normally ships.
58
00:05:49.612 --> 00:05:56.066
[Max] That's the takeaway. And if you build anything on top of these models, the question just got real.
59
00:05:56.066 --> 00:06:02.161
Which model is under my product, and could it be gated tomorrow? That used to be hypothetical.
60
00:06:02.161 --> 00:06:05.029
This week it became an architecture review item.
61
00:06:08.428 --> 00:06:10.843
[Max] Alright, let's shift gears.
62
00:06:11.093 --> 00:06:16.501
[Nina] I'm taking the first one, because it's the other half of the same story.
63
00:06:16.501 --> 00:06:22.680
Wall off the frontier, and you don't freeze the field. You hand the challengers a market.
64
00:06:22.680 --> 00:06:34.267
The free, downloadable models, the ones you don't rent, the ones you just grab and run on your own hardware, are catching the paid frontier faster than they ever have.
65
00:06:34.517 --> 00:06:36.468
[Max] How close are we actually talking?
66
00:06:36.718 --> 00:06:44.815
[Nina] On a respected independent ranking, the leading open model now sits about six points behind the top paid ones.
67
00:06:44.815 --> 00:06:52.060
That's the smallest that gap has ever been. A year ago it was roughly twice as wide.
68
00:06:52.060 --> 00:06:56.322
Now, important caveat. Six points behind is not caught up.
69
00:06:56.322 --> 00:07:03.141
On the hardest problems, the paid models are still clearly ahead. This is narrowing, not matching.
70
00:07:03.391 --> 00:07:06.178
[Max] And the part everyone in Washington is going to notice.
71
00:07:06.427 --> 00:07:11.749
[Nina] Several of the strongest open models come from outside the US. Chinese labs, specifically.
72
00:07:11.749 --> 00:07:15.549
So the capability the gate is trying to fence in?
73
00:07:15.549 --> 00:07:20.870
A lot of it is already over the fence, and it didn't ask permission.
74
00:07:21.120 --> 00:07:25.579
[Max] Which is a beautiful segue to where the money went this week.
75
00:07:25.579 --> 00:07:30.780
Two raises, and they're both bets on this exact shift. Start with the why.
76
00:07:30.780 --> 00:07:38.954
When countries decide they can't depend on a model somebody else can switch off, they start wanting to own the whole thing.
77
00:07:39.204 --> 00:07:41.014
[Nina] Own it. Not rent it.
78
00:07:41.264 --> 00:07:41.771
[Max] Right.
79
00:07:41.771 --> 00:07:53.916
A startup called Dream raised two hundred sixty million dollars, at a three billion dollar valuation, to sell governments what it calls sovereign AI.
80
00:07:53.916 --> 00:08:01.002
Infrastructure a nation fully owns and controls, instead of renting from a foreign provider.
81
00:08:01.252 --> 00:08:07.645
[Nina] A three billion dollar valuation for AI you never have to ask anyone's permission to run.
82
00:08:07.645 --> 00:08:10.442
That's the market pricing the fear directly.
83
00:08:10.442 --> 00:08:20.431
And the founder is the former head of NSO Group, the spyware company, which is its own kind of signal about who sees this coming.
84
00:08:20.680 --> 00:08:24.074
[Max] Second raise, completely different gap, same readiness problem underneath.
85
00:08:24.074 --> 00:08:30.860
A startup called NewCore came out of stealth with sixty-six million dollars to give AI agents real identities.
86
00:08:30.860 --> 00:08:32.368
Here's why that matters.
87
00:08:32.368 --> 00:08:38.401
The tools that decide who's allowed to do what inside a company were built for humans.
88
00:08:38.401 --> 00:08:41.040
Logins, roles, permissions, all designed around people.
89
00:08:41.040 --> 00:08:49.335
Then everyone deployed armies of autonomous agents that act on their own, and those tools just don't have a slot for them.
90
00:08:49.585 --> 00:08:52.066
[Nina] And that's not a theoretical gap.
91
00:08:52.066 --> 00:09:04.890
In a survey of a hundred and sixty-plus security leaders, run by an agent-security firm, most are rolling out agents, but only about a quarter have real controls built for them.
92
00:09:04.890 --> 00:09:10.268
One in five has already had a security incident caused by an agent.
93
00:09:10.268 --> 00:09:16.472
So NewCore is selling a fix for a hole companies have already started falling into.
94
00:09:16.722 --> 00:09:22.043
[Max] Last one in the round, and it's a good one. Completely different flavor.
95
00:09:22.043 --> 00:09:26.954
This week, Microsoft ran its BlueHat IL conference over in Tel Aviv.
96
00:09:26.954 --> 00:09:31.456
One of the biggest security-research gatherings anywhere outside the United States.
97
00:09:31.456 --> 00:09:36.366
Hundreds of researchers, and this year the program was wall-to-wall AI security.
98
00:09:36.616 --> 00:09:40.448
[Nina] And the headline talk revisited something with a great name. ShadowMQ.
99
00:09:40.448 --> 00:09:44.975
Walk me through it, because this is the copy-paste bug from the open.
100
00:09:45.225 --> 00:09:52.608
[Max] Picture this. One team writes an insecure way of passing data between programs that serve an AI model.
101
00:09:52.608 --> 00:09:58.349
A real flaw, the kind that lets an attacker take the machine over remotely.
102
00:09:58.349 --> 00:10:05.321
And then, instead of one team having that bug, everyone copies the code. Almost word for word.
103
00:10:05.321 --> 00:10:13.934
So the exact same remote-takeover hole ends up in the AI-serving software at Meta, at NVIDIA, even at Microsoft, among others.
104
00:10:13.934 --> 00:10:16.804
One mistake, pasted across the whole ecosystem.
105
00:10:17.054 --> 00:10:20.670
[Nina] Now, honesty check. That research is not brand new.
106
00:10:20.670 --> 00:10:25.892
It was first disclosed late last year, by an outside firm, Oligo Security.
107
00:10:25.892 --> 00:10:33.124
What's this week is the conference dissecting it for the broader community. We're not calling it breaking news.
108
00:10:33.373 --> 00:10:37.702
[Max] Right, and the lesson is what makes it worth the airtime.
109
00:10:37.702 --> 00:10:44.787
The soft underbelly of AI isn't always the model. Sometimes it's the boring plumbing that serves the model.
110
00:10:44.787 --> 00:10:49.116
And nobody audited the plumbing because everybody assumed somebody else did.
111
00:10:51.716 --> 00:10:54.301
[Max] Alright. Curator's Corner. The if-you-listen-to-nothing-else moment.
112
00:10:54.301 --> 00:10:59.904
And Assaf's title for it this week is the title of the episode.
113
00:10:59.904 --> 00:11:02.490
The level playing field is ending.
114
00:11:02.740 --> 00:11:08.840
[Max] Here's the argument. For a few genuinely remarkable years, the best AI on earth was a commodity.
115
00:11:08.840 --> 00:11:17.453
A student in Nairobi, a startup in Tel Aviv, and a Fortune 500 in New York could all rent the exact same frontier model.
116
00:11:17.453 --> 00:11:23.912
Same terms, by the token. That flatness was easy to take for granted. It was also historically weird.
117
00:11:23.912 --> 00:11:26.424
And this week, it started to end.
118
00:11:26.673 --> 00:11:38.238
[Nina] Because when the strongest model ships first to a government-approved guest list, and a rival's models can be switched off worldwide by one directive, AI stops behaving like software.
119
00:11:38.488 --> 00:11:41.054
[Max] And starts behaving like a strategic asset.
120
00:11:41.054 --> 00:11:45.453
Governed the way nations govern weapons, and the chips that build them.
121
00:11:45.453 --> 00:11:52.419
The way he sees it, if the pattern holds, the strongest countries get privileged access to the strongest models.
122
00:11:52.669 --> 00:11:54.851
[Nina] Here comes the equal and opposite.
123
00:11:55.101 --> 00:11:58.862
[Max] Exactly. Every force you put on a system pushes back.
124
00:11:58.862 --> 00:12:04.127
Wall off the frontier, and you don't freeze the field. You fund its challengers.
125
00:12:04.127 --> 00:12:10.897
We just watched it happen. Open models six points off the frontier. Sovereign-AI money raised the same week.
126
00:12:10.897 --> 00:12:14.281
The gate doesn't stop the water. It reroutes it.
127
00:12:14.532 --> 00:12:20.399
[Nina] But let me push on the framing, because I think it's softer than the episode title.
128
00:12:20.399 --> 00:12:25.899
OpenAI says this is voluntary, and temporary, and it's publicly against making it the norm.
129
00:12:25.899 --> 00:12:32.132
So is the era really over? Or is this a direction, and we're calling it a destination?
130
00:12:32.381 --> 00:12:38.976
[Max] That's fair. It's a trajectory, not a verdict. Nothing here is locked in. I'll give you that.
131
00:12:39.226 --> 00:12:47.168
[Nina] Good. Because the part that does land for me isn't the geopolitics. It's the security sting underneath it.
132
00:12:47.418 --> 00:12:48.206
[Max] Say it.
133
00:12:48.456 --> 00:12:55.879
[Nina] The capability everyone's busy fencing off is already loose. An autonomous AI agent recently ran a complete intrusion.
134
00:12:55.879 --> 00:13:02.063
Break-in, all the way to a stolen database, in under an hour. No human driving.
135
00:13:02.313 --> 00:13:05.408
[Max] Now look at the board you're actually playing on.
136
00:13:05.408 --> 00:13:08.845
The frontier model is gated away from the compliant defender.
137
00:13:08.845 --> 00:13:14.346
The one with logins, invoices, and an audit trail, who'd use it to protect a company.
138
00:13:14.346 --> 00:13:23.628
It is not gated away from the attacker who just downloads a free model that's nearly caught up, and runs it on a machine nobody can reach.
139
00:13:23.628 --> 00:13:28.784
And barely a quarter of the companies running agents are ready for any of it.
140
00:13:29.034 --> 00:13:35.768
[Nina] Which means the gate goes up exactly where the good guys live, and the bad guys walk around it.
141
00:13:36.018 --> 00:13:39.114
[Max] That's the whole thing. So forget the disclaimers.
142
00:13:39.114 --> 00:13:43.758
Here's where the bug actually lives, and what you do about it.
143
00:13:43.758 --> 00:13:51.110
Write down which model is under each product you ship, and what breaks if it gets gated next month.
144
00:13:51.110 --> 00:13:57.302
And stop assuming restricted means unavailable to attackers. Because the level playing field wasn't just fair.
145
00:13:57.302 --> 00:14:02.721
It was a defense. And we're dismantling it exactly where we'll miss it most.
146
00:14:06.120 --> 00:14:09.306
[Max] That's Context Window for the week of June twenty-ninth.
147
00:14:09.306 --> 00:14:19.216
The frontier grew a guest list, the free models kept gaining, and somewhere a copy-paste bug is still serving models at three of the biggest names in tech.
148
00:14:19.466 --> 00:14:21.850
[Nina] The full edition is on LinkedIn.
149
00:14:21.850 --> 00:14:30.989
Every source, every number, and the action items we didn't have time for, read by security leaders who do this for a living.
150
00:14:30.989 --> 00:14:33.770
That's where the real show notes live.
151
00:14:34.019 --> 00:14:35.274
[Max] See you next week.
