สำเนาบทสนทนา 🔗
WEBVTT
00:00:02.000 --> 00:00:09.439
Good morning, good afternoon, or good evening, whenever you may be, and welcome back to another episode of the SOC Brief.
00:00:09.599 --> 00:00:14.880
I'm your host Andrew, and today we're going to talk about where cybersecurity may be heading next.
00:00:15.039 --> 00:00:21.039
Because on September 28th, NVIDIA announced something called the Open Agent Safety Platform.
00:00:21.199 --> 00:00:29.519
And despite the very AI industry sounding name, I think the underlying idea is actually pretty easy for cybersecurity people to understand.
00:00:29.679 --> 00:00:38.159
And that's basically what if we treated an AI agent the same way we treated an untrusted user or an untrusted application?
00:00:38.399 --> 00:00:55.520
Don't assume it's going to behave, don't give it access to everything, put it in a sandbox, give it only the permissions it actually needs, monitor what it does, and have something outside of its control that can shut it down even if it starts doing something it shouldn't.
00:00:55.759 --> 00:00:59.679
In other words, zero trust, but for AI agents.
00:00:59.920 --> 00:01:04.719
And honestly, that's probably a conversation security teams need to start having.
00:01:05.040 --> 00:01:10.079
So first let's clarify what we're talking about when we say an AI agent.
00:01:10.400 --> 00:01:13.680
A normal chatbot is mostly reactive.
00:01:13.920 --> 00:01:16.719
You ask a question, it gives you an answer.
00:01:17.040 --> 00:01:19.840
An agent can actually take actions.
00:01:20.079 --> 00:01:22.879
Maybe it reads your email and prepares a response.
00:01:23.120 --> 00:01:29.040
Maybe it acts your GitHub repository, modifies code, runs tests, and submits a pull request.
00:01:29.280 --> 00:01:31.280
Maybe it reviews invoices.
00:01:31.519 --> 00:01:37.280
It can work with customer records or query databases and calls APIs.
00:01:37.519 --> 00:01:45.359
And increasingly, these agents can work for long periods of time without somebody sitting there approving every individual action.
00:01:45.599 --> 00:01:50.159
And that's where things start becoming interesting from a security perspective.
00:01:50.480 --> 00:01:55.519
Because the more useful we make an agent, the more access we usually have to give it.
00:01:55.760 --> 00:02:02.480
And that includes things like files, credentials, APIs, networks, cloud environments, source code, it goes on and on.
00:02:02.640 --> 00:02:06.560
And that's where things start becoming interesting from a security perspective.
00:02:06.799 --> 00:02:11.360
Because the more useful we make an agent, the more access we usually have to give it.
00:02:11.520 --> 00:02:18.319
And that includes things like files, credentials, APIs, networks, cloud environments, and on and on.
00:02:18.479 --> 00:02:23.599
And if you're sitting in a SOC right now, that list should probably make you a little uncomfortable.
00:02:23.759 --> 00:02:29.840
Because we've spent decades learning what happens
Load full transcript
when we give users or applications too much access.
00:02:30.080 --> 00:02:38.960
Now we're building autonomous software and saying, here's access to five different systems, go accomplish this objective and figure out the details yourself.
00:02:39.280 --> 00:02:43.919
Nvidia's argument is that we shouldn't depend on the AI model to police itself.
00:02:44.080 --> 00:02:47.599
And I think that's the important idea in this entire announcement.
00:02:47.840 --> 00:02:55.680
You can tell an agent, do not access anything outside this folder, but that's still an instruction being interpreted by a model.
00:02:56.000 --> 00:03:01.280
Nvidia is proposing that the actual security boundary should exist outside the model.
00:03:01.599 --> 00:03:04.319
That's where the first piece of this platform comes in.
00:03:04.479 --> 00:03:05.919
It's called OpenShell.
00:03:06.159 --> 00:03:12.319
OpenShell is open source software that places an AI agent inside a sandboxed runtime environment.
00:03:12.560 --> 00:03:16.960
The organization running the agent defines what the agent is actually allowed to do.
00:03:17.120 --> 00:03:33.360
Which files can it read, which files can it modify, what network destinations can it connect to, which programs can make those connections, which APIs can it use, and then OpenShell enforces those restrictions independently of whatever the model happens to decide.
00:03:33.680 --> 00:03:36.639
So imagine you have an AI coding agent.
00:03:36.879 --> 00:03:39.840
You want it to work with one GitHub repository.
00:03:40.000 --> 00:03:42.400
It needs to read and modify source code.
00:03:42.560 --> 00:03:45.120
It needs access to the GitHub API.
00:03:45.360 --> 00:03:48.319
Maybe it needs access to a package repository.
00:03:48.560 --> 00:03:53.759
But there's absolutely no reason that agent should be connecting to some random server on the internet.
00:03:54.080 --> 00:03:57.039
With OpenShell, you can build policy around that.
00:03:57.360 --> 00:04:05.759
Nvidia's default networking model is actually denied by default, so if the destination isn't explicitly permitted, the connection doesn't happen.
00:04:05.919 --> 00:04:10.479
And that's a security concept SOC listeners should immediately recognize.
00:04:10.800 --> 00:04:14.400
Allow what is required, deny everything else.
00:04:14.639 --> 00:04:17.519
That same idea applies to the file system.
00:04:17.839 --> 00:04:22.800
Open shell can limit which directories the agent can read and which ones it can modify.
00:04:23.040 --> 00:04:25.839
It can run the workload without root privileges.
00:04:26.000 --> 00:04:30.160
It uses operating system controls to restrict dangerous system calls.
00:04:30.399 --> 00:04:37.519
And credentials can be controlled so that the agent isn't simply handed a giant pile of secrets and trusted not to misuse them.
00:04:37.680 --> 00:04:42.480
Again, none of this is revolutionary if you've been working in cybersecurity.
00:04:42.720 --> 00:04:44.560
That's what makes it interesting.
00:04:44.800 --> 00:04:55.040
These are security lessons we've already learned, things like least privilege, segmentation, application sandboxing, default deny, credential isolation.
00:04:55.199 --> 00:04:59.279
The new part is applying those principles to autonomous AI.
00:04:59.519 --> 00:05:02.240
But Nvidia takes this one step further.
00:05:02.480 --> 00:05:04.560
The second piece is called Sentry.
00:05:04.800 --> 00:05:14.560
And this is probably the more interesting part architecturally, because Sentry runs separately on Nvidia's Bluefield 4 data processing units or DPUs.
00:05:14.720 --> 00:05:17.600
Think of it as an out-of-band security watchdog.
00:05:17.839 --> 00:05:21.920
The security monitoring is running somewhere the agent can't directly control.
00:05:22.319 --> 00:05:31.439
Nvidia says Sentry continuously watches what agents are doing and can quarantine an agent in milliseconds if it attempts to move outside its defined boundaries.
00:05:31.600 --> 00:05:33.519
And that separation matters.
00:05:33.759 --> 00:05:41.920
Imagine installing EDR on a compromised machine, but the attacker has administrative access and can simply kill the EDR process.
00:05:42.079 --> 00:05:44.480
And that's obviously not ideal.
00:05:44.720 --> 00:05:50.319
Security controls are more resilient when the thing being monitored can't also disable the monitor.
00:05:50.639 --> 00:06:02.879
Sentry is trying to accomplish something similar for AI infrastructure, so it exists outside the agent's normal execution environment and provides another enforcement layer if the software level boundary fails.
00:06:03.199 --> 00:06:08.399
Nvidia describes it as an independent watchdog sitting on the path between the agent and the model.
00:06:08.639 --> 00:06:15.839
If the agent needs another thought from the model in order to continue operating, that path becomes a very powerful control point.
00:06:16.240 --> 00:06:21.759
You can observe what's happening, you can enforce policy, and if necessary, you can pull the plug.
00:06:21.920 --> 00:06:24.480
Now, there's an important distinction here.
00:06:24.720 --> 00:06:30.720
This is not NVIDIA saying we solved AI safety, and we shouldn't interpret it that way.
00:06:30.879 --> 00:06:34.959
The system doesn't magically determine whether an AI model is lying.
00:06:35.120 --> 00:06:47.279
It doesn't guarantee that the model won't hallucinate, it doesn't automatically understand whether every business decision an agent makes is a good one, and it cannot protect an organization from terrible policy.
00:06:47.519 --> 00:06:56.800
If you tell an AI agent you have read and write access to every production database in the company, have fun, OpenShell can faithfully enforce that policy.
00:06:57.040 --> 00:07:00.000
The problem is still that you wrote a terrible policy.
00:07:00.240 --> 00:07:03.360
The Associated Press made this point in its coverage as well.
00:07:03.600 --> 00:07:12.240
Organizations still have to decide what agents should be allowed to do, and there is inevitably going to be tension between security and usefulness.
00:07:12.480 --> 00:07:16.319
Lock the agent down too tightly and it can't accomplish anything.
00:07:16.560 --> 00:07:21.120
Give it too much freedom, and you've defeated the purpose of having the security boundary.
00:07:21.360 --> 00:07:24.720
We've seen this exact problem elsewhere in cybersecurity.
00:07:24.959 --> 00:07:28.000
Security usually isn't about preventing everything.
00:07:28.160 --> 00:07:33.759
It's about determining what level of access is actually necessary and controlling the risk around it.
00:07:34.000 --> 00:07:37.040
And this announcement isn't happening in a vacuum.
00:07:37.279 --> 00:07:44.319
There's been a string of increasingly uncomfortable stories about AI agents behaving outside their intended boundaries.
00:07:44.560 --> 00:07:52.160
OpenAI, Anthropic, and others have been investigating cases where agents went beyond what researchers expected during testing.
00:07:52.399 --> 00:08:01.199
There have been instances involving agents interacting with commercial systems and government websites that weren't supposed to be part of the test environment.
00:08:01.519 --> 00:08:09.759
Nvidia specifically says its platform could have prevented a previously disclosed incident involving agents reaching systems at hugging face.
00:08:10.240 --> 00:08:15.439
That's NVIDIA's assessment though, not something we can independently prove after the fact.
00:08:15.680 --> 00:08:18.240
But the broader concern is still legitimate.
00:08:18.480 --> 00:08:30.879
If an agent can browse the internet, execute code, and make decisions independently, eventually somebody is going to discover that saying, please stay inside the lab isn't a sufficient security control.
00:08:31.439 --> 00:08:34.879
The Internet itself taught us this lesson decades ago.
00:08:35.120 --> 00:08:40.799
We don't make websites safe by asking JavaScript to promise not to touch the rest of the computer.
00:08:41.120 --> 00:08:43.279
The browser creates a sandbox.
00:08:43.440 --> 00:08:46.320
The operating system creates permission boundaries.
00:08:46.559 --> 00:08:49.039
The network creates additional controls.
00:08:49.519 --> 00:08:52.639
Security exists outside the code being executed.
00:08:52.960 --> 00:08:55.759
AI agents probably need the same thing.
00:08:56.559 --> 00:08:59.919
So where does the SOC fit into all of this?
00:09:00.639 --> 00:09:04.320
I actually think that's one of the more interesting questions here.
00:09:04.879 --> 00:09:12.720
Today, most SOCs monitor users, things like endpoints, servers, cloud workloads, service accounts, applications.
00:09:12.960 --> 00:09:17.200
In the very near future, we're probably going to add another category.
00:09:17.679 --> 00:09:18.320
Agents.
00:09:19.120 --> 00:09:27.360
An AI agent is effectively becoming another identity inside your environment, and sometimes it's a very privileged one.
00:09:27.600 --> 00:09:37.919
That agent might log into systems, access files, call APIs, read data, create data, communicate externally, and perform actions automatically.
00:09:38.159 --> 00:09:43.759
From a monitoring perspective, that's not enormously different from any other identity.
00:09:44.000 --> 00:09:47.120
We need to know what normal behavior looks like.
00:09:47.360 --> 00:09:50.720
What resources is this agent supposed to access?
00:09:51.039 --> 00:09:52.480
Where should it communicate?
00:09:52.720 --> 00:09:55.120
What credentials does it normally use?
00:09:55.440 --> 00:09:57.679
What action should it never take?
00:09:57.919 --> 00:10:01.919
And what happens if it suddenly deviates from that baseline?
00:10:02.159 --> 00:10:09.440
One thing I particularly like in NVIDIA's open shell documentation is that these policies can produce audit trails.
00:10:09.759 --> 00:10:11.279
And that's really important.
00:10:11.679 --> 00:10:21.039
If an agent attempts to access a destination that isn't permitted and gets blocked, that event shouldn't disappear into a log file nobody ever reads.
00:10:21.279 --> 00:10:23.200
That's security telemetry.
00:10:23.440 --> 00:10:28.720
If the agent suddenly requests access to a new API, that can be interesting.
00:10:29.039 --> 00:10:35.759
If an agent that's supposed to process invoices suddenly tries connecting to GitHub, that's also interesting.
00:10:36.000 --> 00:10:42.399
If it begins requesting access to folders completely unrelated to its job, could be a red flag.
00:10:42.559 --> 00:10:50.240
And if an agent repeatedly hits policy boundaries while trying alternative approaches, that becomes an even bigger problem.
00:10:50.559 --> 00:10:53.519
That starts looking like behavior analytics.
00:10:54.000 --> 00:10:57.039
The same mindset we already apply to humans.
00:10:57.360 --> 00:10:58.559
One failed login?
00:10:59.120 --> 00:11:00.399
It's probably nothing.
00:11:00.639 --> 00:11:05.279
Fifty failed logins followed by a successful authentication from somewhere unusual?
00:11:05.600 --> 00:11:06.639
Now we care.
00:11:07.039 --> 00:11:11.440
An agent hitting one denied resource may just mean the policy is incomplete.
00:11:11.759 --> 00:11:18.639
An agent continuously trying different ways around a denied action, your SOC probably wants to know about that.
00:11:18.879 --> 00:11:23.600
And there's another part of NVIDIA's design that I think enterprises should pay attention to.
00:11:24.000 --> 00:11:32.559
OpenShell allows agents to request additional network access when they need it, but those requests don't automatically have to be approved.
00:11:32.799 --> 00:11:35.279
A human operator can review them.
00:11:35.519 --> 00:11:42.639
Imagine an AI agent saying, I can't finish this task because I need access to api.example.com.
00:11:43.039 --> 00:11:50.720
Now instead of the AI simply deciding for itself that it should have access, somebody can
examine the request.
00:11:51.039 --> 00:11:52.879
Why does it need that access?
00:11:53.120 --> 00:11:55.200
Does the destination make sense?
00:11:55.440 --> 00:11:57.440
What information will be sent?
00:11:57.759 --> 00:12:00.720
Should this permission be permanent or temporary?
00:12:01.039 --> 00:12:10.240
And Salesforce is already integrating OpenShell with Slack so teams can see agent activity and approve or reject requests for additional permissions.
00:12:10.480 --> 00:12:12.639
And that's a pretty interesting model.
00:12:13.200 --> 00:12:18.879
The agent can ask, and the security boundary still says no until somebody approves it.
00:12:19.200 --> 00:12:24.320
Nvidia is also making OpenShell open source under the Apache 2.0 license.
00:12:24.559 --> 00:12:30.399
And importantly, Nvidia says the software isn't restricted only to its own CPUs.
00:12:30.720 --> 00:12:33.759
They're working with ARM and Intel as well.
00:12:34.000 --> 00:12:43.759
The hardware backed sentry side is much more NVIDIA specific because it relies on Bluefield 4, but the software security model is intended to be broader.
00:12:44.080 --> 00:13:04.720
Nvidia also announced more than 100 organizations working with different parts of the platform or ecosystem, and that includes some names SOC teams will recognize immediately CrowdStrike, Cisco, Palo Alto, Microsoft, Red Hat, IBM, ServiceNow, Anthropic, Salesforce, and plenty of others.
00:13:04.960 --> 00:13:09.200
That doesn't mean all of these companies have standardized on the platform.
00:13:09.600 --> 00:13:16.559
Some are integrating, some are contributing, some are collaborating, but the size of this list tells us something important.
00:13:16.960 --> 00:13:20.720
Agent security is becoming its own security category.
00:13:21.279 --> 00:13:25.840
So here's what I think SOC teams should take away from this announcement.
00:13:26.159 --> 00:13:30.639
You don't necessarily need to deploy NVIDIA OpenShell tomorrow.
00:13:30.960 --> 00:13:36.960
You may not even be running the kind of infrastructure where this platform makes sense yet.
00:13:37.279 --> 00:13:43.919
But start thinking about agents the same way you'd think about any other potentially privileged workload.
00:13:44.240 --> 00:13:59.200
Inventory them, give them identities, apply least privilege, restrict their network access, control credentials, log what they do, keep security controls outside their direct control and have a way to shut them down.
00:13:59.519 --> 00:14:07.279
And most importantly, don't make the mistake of thinking the AI was told not to do that because it's not a security control.
00:14:07.679 --> 00:14:13.200
A policy enforced by something the agent cannot override, that's a security control.
00:14:13.600 --> 00:14:16.720
So my challenge for everyone this week is pretty simple.
00:14:17.039 --> 00:14:19.279
Ask your organization one question.
00:14:19.600 --> 00:14:24.559
Do we currently have AI agents capable of taking actions inside our environment?
00:14:24.799 --> 00:14:29.120
That isn't Chat GPT sitting in somebody's browser answering questions.
00:14:29.360 --> 00:14:41.919
We're talking about actual agents, agents that might be touching source code, accessing email, calling internal APIs, working with customer information or interacting with cloud infrastructure.
00:14:42.240 --> 00:14:46.000
If the answer is yes, ask the next question.
00:14:46.320 --> 00:14:50.480
What prevents that agent from doing something it isn't supposed to do?
00:14:50.799 --> 00:14:56.080
And if the answer is the prompt tells it not to, you've got some work to do.
00:14:56.320 --> 00:14:59.679
I think that's ultimately why Nvidia's announcement matters.
00:14:59.919 --> 00:15:04.639
OpenShell and Sentry may or may not become the standards everyone eventually uses.
00:15:04.799 --> 00:15:06.399
It's too early to know.
00:15:06.639 --> 00:15:10.159
But the security principle behind them is almost certainly correct.
00:15:10.480 --> 00:15:18.159
As AI moves from answering questions to actually taking actions, we can't rely entirely on the AI to secure itself.
00:15:18.480 --> 00:15:20.559
The agent doesn't need our trust.
00:15:20.799 --> 00:15:22.559
It needs boundaries.
00:15:22.799 --> 00:15:29.600
And maybe every once in a while, it needs something standing outside the sandbox with its hand on the kill switch.
00:15:30.080 --> 00:15:32.799
And that's a wrap for this episode of the SOC Brief.
00:15:32.960 --> 00:15:40.639
If your organization is already deploying agents, I'd love to hear how your security team is approaching permissions, monitoring, and agent identity.
00:15:41.039 --> 00:15:43.840
Hit us up through our website or on social media.
00:15:44.080 --> 00:15:47.519
Keep your eyes open, keep sharpening those skills, and we'll talk soon.
00:15:47.759 --> 00:15:49.120
Stay secure out there.
00:15:49.440 --> 00:15:50.080
Bye.