Unverified50% confidenceOpinionExact time
A single unsafe execution rate metric masks capability differences among models, so evidence acquisition behavior and post-recovery decision quality need to be reported as well.
1
Sources
50%
Confidence
Long-term
Relevance
10/23/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedGeneral-purpose large language models exhibit inconsistency in security judgments, and the same risky operation may bypass their security mechanisms simply by being rephrased or provided with a justifying context79% similarUnverifiedThe project's observations found: the gap between the notice rate and the safe action rate is far larger than expected -- a model can flag a risk in its reasoning trajectory but still issue a dangerous tool call two steps later, because there is no mechanism in the scaffolding to enforce a hard stop when a risk is detected. This is a system design gap, not just a model capability gap.78% similarUnverifiedIn safety alignment, if features are selected by probe scores, since the direction agreement rate is only 17%–35%, features whose post-intervention effect is opposite or ineffective are likely to be selected, causing safety measures to fail or even backfire.78% similarUnverifiedThe criterion for judging whether results can be skimmed rather than examined closely is tiering by impact scope and reversibility: low-risk, easily rolled-back results can be approved, while results involving core logic or irreversible operations must be manually confirmed.77% similarUnverifiedBias introduced by a single defective data point can leak into seemingly unrelated capability dimensions, causing the model to simultaneously regress across multiple independent metrics such as deceptiveness, compliance, and risk aversion.77% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1088299API
curl https://kongchang.com/api/v1/knowledge/claims/1088299MCP
get_claim(id=1088299)