Why an Agent That Learned to Be Safe Somewhere Else Isn't Safe on Your Unit

A question the market is asking after a summer of agent headlines: if an AI agent learned to behave safely on the tasks it was trained on, will it behave safely on a task it has never seen?

Our CTO Nadav Cohen's group at Tel Aviv University answered it in a paper this spring, and the answer is no, and not because of a training shortcut. They proved that safe execution of a task is mathematically more sensitive to what the task is than execution alone. Doing the job transfers. Doing it safely does not.

They showed it three ways: in a classical control problem, in a simulated quadcopter, and in a language-model agent working a CRM system.

The consequence for a plant is direct. A general-purpose agent that was safe somewhere else is not, by that fact, safe on your unit.

We built the other way. A closed-loop DLPC® is built per unit, runs inside bounds the operator sets, and is supervised. The safety is in the design, not in the hope that training carried over.

Read the paper: "Why Does Agentic Safety Fail to Generalize Across Tasks?" (Slutzky, Alexander, Slor, Nagel, Cohen)