Kensa検査
Kensa provides runtime guardrails and a security system for AI agents, built on the model’s internal state. It watches every token at the level of neural activations as the agent thinks, and stops an unsafe action before the command is even finished.
The problem
Guardrails watch the ends. Agents fail in the middle.
Why reading the text is not enough
What a model reveals is the surface. What it computes—and acts on—runs deeper.
Inspecting chain-of-thought isn’t enough: it only reacts after intent surfaces, and only if the model chooses to show it. Kensa reads the computation underneath, intercepting unsafe intent where it forms first.
How often reasoning models reveal a hint they actually used, in many settings.Chen et al., Anthropic, 2025
Agreement between what a model reveals in its chain-of-thought and its internal reasoning.Hong et al., 2026
More reward hacking when an internal “desperation” state is amplified, with no clear sign in the chain-of-thought.Sofroniew et al., Anthropic, 2026
How it works
Every generation is a walk through named states.
Illustrative data shaped like a real hijack.
Runtime guardrails
Layered defence, with a sensor no text filter has.
Many agents run at once, and threats arrive at every stage. Each layer catches a different kind: injected prompts at the door, poisoned content as it is read, unsafe calls before they run, leaks on the way out. The hijacks that leave no trace in the text are caught from the inside. Select a stage to see its guard.
Research