Unverified50% confidenceOpinionExact time
Usage reports are only a signal for investigation, not proof that the agent followed every instruction, nor proof that a certain error would necessarily occur without the rule.
1
Sources
50%
Confidence
Long-term
Relevance
10/13/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedRecording only the proposed calls and tool responses is insufficient to determine the true cause of an agent security test result.74% similarUnverifiedA model's self-report cannot be used to detect whether it followed prohibited information; runtime auditing and external behavioral monitoring are more reliable than a model's introspective statements.74% similarUnverifiedSigned evidence can only prove that the report has not been tampered with, and cannot prove that the agent itself is safe.73% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1020452API
curl https://kongchang.com/api/v1/knowledge/claims/1020452MCP
get_claim(id=1020452)