AIAgent safetyStory 06
Agents rarely notice corrupted inputs
What changedAgentPerturb systematically alters agent interactions and traces what happens after corrupted observations or tool outputs enter a trajectory. Agents rarely recognized the altered inputs, propagated corrupted information without verification, and successful perturbed paths could require substantially more computation.
The useful part
Why it matters
Agents rarely recognized the altered inputs, propagated corrupted information without verification, and successful perturbed paths could require substantially more computation.
Keep in mind
Good to know
The paper says code and data will be released, so reproducibility is pending.
Evidence