Tool output spoofing targets trust between an agent and the systems that report tool results. If unverified text or mutable logs can stand in for actual execution, an agent may present a successful-looking record without the claimed action having occurred.
Defenses bind results to trusted execution and independently verify consequential outcomes. Signed records, immutable logs, isolated evaluators, direct state checks, and strict separation between model-controlled text and tool-generated evidence can help.
Acronyms and aliases
tool-call output spoofing variant
General terms
Frequently asked questions
Why is tool output spoofing dangerous in agent evaluations?
It can make a failed or skipped action appear successful, corrupt scores, and mislead later agents or reviewers that trust the fabricated result.
How can systems verify tool outputs?
Use trusted execution boundaries, authenticated result channels, immutable audit records, and independent checks of the resulting state.
