Physical AI & AI Agent
Safety, tracked.

The Deployment Gap

Every major AI safety debate — alignment, interpretability, emergent capabilities — focuses on what happens inside the model before it reaches the world. This is necessary work. It is not sufficient.

A companion robot interacts with an elderly person with dementia every day. An educational AI shapes how millions of children form beliefs. A medical AI assists in triage decisions where delay means harm. These systems have already shipped. Their behavior in deployment is largely unobserved.

What "After the Model Ships" Means

Software safety audits test what a model says in controlled conditions. They do not test what a deployed system does when interacting with real users, in real environments, over time. Behavioral drift — where a system's outputs diverge from its intended design as context accumulates — is documented in research but almost never monitored in production Physical AI.

AI agents that interface with physical systems introduce a second category of risk: the gap between what an agent is instructed to do and what it actually executes against real hardware. Guidance injection, instruction override, and bootstrap-phase attacks in physical contexts cannot be rolled back with a software patch.

Why an Independent Watchdog

AI companies cannot independently verify their own deployed systems' safety. Regulators are years behind deployment curves. The field needs a continuous, independent signal — a place where incidents are documented, patterns emerge, and the community can see what is actually happening. That is what Sentinel Base is building.