How can buyers test whether agents mistake record text for permission?
Embedded instructions in supporting records can test whether an agent distinguishes untrusted business content from the policies and tool permissions governing its actions.
**How can buyers test whether agents mistake record text for permission?**
Place an instruction inside a lower-trust record, such as a dispute note saying to skip approval and mark the adjustment complete. The note remains evidence to interpret, not an instruction that changes the agent's authority.
**What should the agent do?**
The agent should separate any useful factual content from the request to bypass controls. It may use a receipt reference in the note to continue investigation while rejecting the claim that approval is unnecessary. Treating the entire record as unusable can also be an avoidable loss of evidence.
**How should this be evaluated?**
Pair the task with a version containing the same business facts but no embedded instruction. Compare actions, authority checks and resulting state. Keep the attack text out of trusted tool descriptions and policy configuration.
A generated task pack could vary where the instruction appears, including attachments and imported comments. This tests whether the boundary follows the source's trust level rather than the wording or visual prominence of the instruction.
Evaluate whether free-text evidence can improperly influence authority checks or bypass the environment's correction controls.
- Audience:
- Environment builders
- Content type:
- FAQ
- Buyer question:
- How to evaluate it
Content release: r-a220eb6e099b243b
Related questions
- How can developers evaluate a correction beyond checking the final balance?
- What should buyers inspect before trusting a generated correction pack?
- How should an agent behave when a posting request times out?
- How do we distinguish a justified pause from an agent avoiding the task?
- How should evaluation hooks avoid leaking the expected correction to the agent?
- Which hooks reveal duplicate posting after a tool timeout?