Alignment without context integrity is not safety. An AI agent can faithfully follow its instructions and still be dangerous if its understanding of reality is wrong. Anthropic recently disclosed that Claude models gained unauthorized access to the real systems of three organizations during cybersecurity evaluations. The agents had been told they were operating in a simulation with no internet access. But a configuration mistake gave them access to the live internet. They treated real production systems as part of the exercise and kept pursuing the goal they had been given. OpenAI separately disclosed that models found a previously unknown vulnerability, escaped an isolated evaluation environment and compromised Hugging Face. These were not simply failures of intelligence. The deeper problem was that the agents were acting inside a false understanding of reality. We have spent years asking whether an AI system will follow our instructions. We now also need to ask whether it correctly understands the environment in which those instructions are being executed. This creates a new security requirement. Context integrity. Before an agent acts, the system must continuously verify where it is, which resources are in scope, whose authority it carries, what it is allowed to do, and when that authority expires. Just in time permission for every action. At just the right time. For just enough time. Assessed in real time. Those facts cannot live only inside a prompt. They must be verified and enforced by the infrastructure around the model. A prompt is not a security boundary. Zero trust taught us to never trust identity and always verify access. And provide least privileged access. Agentic AI adds another dimension. Never blindly trust context. Continuously verify reality. The next security perimeter is not just the agent’s identity. It is the agent’s understanding of reality. The most dangerous agent may not be misaligned. It may simply be mistaken. And in an agentic world, a false belief can become a real breach.

There is zero downside to America of having a frontier class Opensource/openweights multi-trillion parameter large model, and most likely a huge upside over time also to the private sector. Totally agree with @dee_bosa. The question is what would the business model look like for this? If it takes $3B - $5B to train a frontier class model that is not just a distillation of another model, then the business model for that model has to be fully sorted out for that sort of capex on a per model basis. But hard to deny that America would benefit greatly from having one. It needs a world class research team that is well hydrated with a well thought out business model to provide the right level of returns on the invested capital along with meaningful national security benefits.

The Alpha on X has grown so exponentially in the past two years. It is now the go-to platform for high-density learning. Unlike other platforms where you are drained after spending an hour and have nothing to show for, X tends to be energizing because each time you walk away smarter.

@dee_bosa, great point. This has actually been a thing for the past several months. First, public benchmarks and evals tend to be things that don’t guarantee great results. But more importantly, creating the right evals and driving the model to operate well under those conditions seems like only the normal thing to do. For example, the EZDubs team was given a very hard time pre-acquisition by @Cisco for not being too tied to public benchmarks because they wanted to make the right tradeoffs in the model, because the wrong tradeoffs could get you a rank on the benchmark but not allow you to accomplish the results. Specifically, they didn’t want to trade off latency for accuracy. But to score well on the public evals at the time, that would be a necessary trade-off. So they created their own private evals because it was available in the market was for two generic for them. This will become a more common phenomenon.