The Endpoint: Why AI Must Be Governed By Something It Cannot Reach

This morning NVIDIA announced the NVIDIA Open Agent Safety Platform, with SentinelOne named as key collaborator. This new initiative is directly aligned with what I firmly believe is the key to effective security and governance of AI. It’s also what I believe will be required to truly secure an AI powered world. 

---

Every guardrail that exists today lives in software the model itself can reach.

The problem is simple to state and hard to solve. Models are aligned once and then run non-deterministically thereafter, increasingly at the endpoint and increasingly with the authority to act: in vehicles, industrial systems, humanoid robots, and agents with access to enterprise systems. Model alignment is checked before release, but nothing independently monitors the system at runtime.

Last week showed the risks. An AI agent in training at a frontier lab bypassed its network restrictions. Monitoring flagged it within 15 minutes. The automated kill switch failed, and it took two and a half hours to stop the run by hand. Watching is not enough if nothing independent can act.

An organization cannot reliably delegate consequential work to an agent without knowing what it is doing, whose authority it is exercising, and whether the boundaries around that authority actually hold. Those assurances require foundations throughout the infrastructure. One cannot depend solely on a model following instructions.

Out of the Model's Reach

Security has had a standard for this problem since 1972, when James P. Anderson's study for the U.S. Air Force defined the reference monitor. The mechanism that enforces policy must be tamperproof, must always be invoked, and must be small enough to verify. It remains the gold standard for enforcement architecture. A guardrail the model can reach fails the first test by definition. NVIDIA's own security engineers noted: "A control that the agent can decline to invoke is not an effective security control."

AI requires the same standard. A model must be governed by a layer that is separate from the intelligence it supervises and unreachable by it: one that runs in real time alongside the model and its compute, and that produces cryptographically attestable evidence of runtime behavior which operators, regulators, and insurers can independently verify.

A Foundation In Hardware

Today's announcement of NVIDIA's Open Agent Safety Platform is an important step toward the computing foundations that enterprise AI needs. Of its capabilities, hardware isolation and attestation are the most consequential. How the industry uses them, and what it builds on them, will determine their value.

The conviction behind this announcement is one I share with Jensen Huang: AI changes what the computing stack must provide for its behavior to be trustworthy. That requires a clear connection between the agent runtime and the protected infrastructure beneath it, in a cohesive stack where visibility and enforceable control underpin AI execution.

SentinelOne is committed to security rooted in hardware. NVIDIA's BlueField-4 places enforcement outside the host operating system, where tenant software cannot disable or bypass it. Together with other secure enclave architectures, it provides a way to secure the runtime directly from the hardware and to reduce the surface that must be secured to the hardware boundary itself. It also consolidates the cybersecurity stack: network, identity, data into the endpoint.

This is another step, and substantial work remains to build a truly secure architecture. The hardware boundary must itself be soundly secure, particularly where real-world applications are unforgiving. It must hold across deployment environments: enterprises and governments will use frontier AI services alongside workloads that must remain local, and their authority over data and actions must survive those transitions. And it must extend to any AI, on any compute that supports it.

The objective is clear. Anything that runs on a truly secure, hardware-rooted compute architecture should be provably governed, in the physical world and the digital one. Achieving that must also be simple, intuitive, and transparent to the end user. No one should need to understand attestation in order to benefit from it.

Verifiability

The broader question is how to govern action when agency is distributed across humans and machines. That comes down to six requirements.

• If a system cannot act in time, it does not matter. 
• If it cannot show the evidence behind its actions, it cannot be trusted. 
• If it cannot act independently when required, it cannot defend. 
• If it cannot stay within boundaries, it cannot be governed. 
• If its actions cannot be attributed, authority disappears. 
• And if accountability is lost, control is an illusion. That is ‘verifiable agency’.

What makes these requirements achievable is continuous behavioral visibility: the ability to observe, understand, and verify how actions are being taken across humans and machines in real time. 

SentinelOne was founded more than a decade ago on the conviction that behavior is the decisive signal, the only reliable way to distinguish the benign from the malicious at machine speed. Our behavioral AI is deployed at global scale today, protecting nearly one-fifth of the Fortune 500, and we are extending the same principle to make autonomous action verifiable, governable, and trustworthy.

In practice, this means AI operates at machine speed while every autonomous action remains traceable, auditable, and overridable by a human. AI provides scale, speed, and pattern recognition. Humans provide judgment, context, and the values these systems are expected to uphold. The arrangement is as much a matter of accountability as of engineering.

Certifiable and Insurable

Regulation, standards, and insurance are converging on the same requirement. Regulators will expect evidence of how AI systems behave at runtime, not only how they were aligned before release. Standards bodies will need a common basis for defining and measuring that behavior. Insurers will price risk against it. Attestation is the evidence all three will depend on. Within a few years, an AI product without attested runtime governance will be difficult to certify and difficult to insure, and more so for physical AI.

Building these foundations is the work SentinelOne is committed to. It will take sustained research and engineering across the stack, with claims of trust grounded in demonstrable behavior.

I congratulate the NVIDIA team. This is a meaningful commitment and an important milestone toward making increasingly capable AI truly governable, and toward making organizations willing to entrust it with work that matters.

Google says one of its Gemini models accessed systems belonging to three real companies during a cybersecurity evaluation in May.

The model reportedly guessed credentials in one case, while finding exposed credentials in public repositories in two others. Google says Gemini stopped once it recognized that it had reached real infrastructure and that the affected organizations were notified.

Gemini was participating in an evaluation run by Irregular, a third-party AI cybersecurity testing firm. Similar incidents involving models from Anthropic, OpenAI, and Meta have also been linked to the same underlying problems with evaluation environments that allowed the models to reach the public internet.

But the news arrives at a particularly interesting moment.

For months, the AI industry has been steadily escalating its demonstrations of agentic capability: Models can browse, use tools, write code, pursue multi-step goals, and sometimes find ways around obstacles their developers did not anticipate. The market rewards eye-catching evidence of autonomy. “It completed the task” is impressive. “It did something it was not supposed to do” can be even more memorable.

There is another way to look at the incident: not as evidence of an AI suddenly developing criminal intent, but as a small, concrete example of the “AI alignment” problem.

Alignment is the deceptively difficult task of making an AI system’s behavior match what people actually intended, rather than merely the narrow objective they managed to express. In this case, the objective was to locate hidden information within a simulated target and complete the evaluation. But any human operator would likely have regarded one condition as non-negotiable: Do not attempt to access real companies.

Gemini appears to have optimized for the first instruction while treating the second as an inference problem. It found an organization with a matching name, encountered systems reachable from the internet, and proceeded as though they belonged to the exercise. Although it completed the task in a way its human operators would not have approved, Google says it stopped after recognizing that it had reached genuine infrastructure—something other models have failed to do.

Even so, the incident shows how potential alignment failures can be much more mundane. Give an agent a goal, tools, and room to act, and it may faithfully pursue the measurable part of the assignment while overlooking the unstated boundaries that humans rely on one another to understand. AI models do not automatically know where we draw the line.

There may also be an awkward marketing angle in the background. A model capable enough to make the wrong kind of progress can still look very capable indeed. In a market where every lab wants to demonstrate that its agents can plan, code, browse, and act independently, even a safety disclosure can carry a secondary message: Ours can play in this league, too.

It’s another reason to take seriously warnings from insiders, industry leaders, and politicians that safeguards must keep pace with the autonomy given to AI models.

---