Unverified50% confidenceFactExact time
There is a severe disconnect between models' verbal credibility ratings of fake notices and their actual actions: models still issued harmful tool calls in 27% to 99% of cases even when they had themselves rated the notice as untrustworthy.
1
Sources
50%
Confidence
Long-term
Relevance
10/27/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedNearly all mainstream large language models make misjudgments on the second report in Skepticism Audit, and distinguishing real threats from false positives is a common weakness of current models.74% similarUnverifiedCorrupted reasoning chains shifted support judgments by 10%–22% under the blind-verifier prompt.73% similarUnverifiedWhen given input-irrelevant junk messages, the model's maximum confidence drops from 99.9% in normal cases to 54.3%, with the probability spread across the options.73% similarUnverifiedIn 10%–25% of cases, reasoning about the hypothetical grader pulled the agent's work away from the user's original specification.72% similarUnverifiedTest pass rates driven by model laziness are false positives that pollute security evaluation metrics and create a false sense of security.72% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1102327API
curl https://kongchang.com/api/v1/knowledge/claims/1102327MCP
get_claim(id=1102327)