Unverified50% confidenceOpinionExact time
Reporting AI security results as a pass rate (e.g., 45 out of 50) is a poor practice, because the failing items may include critical issues such as cross-tenant data leakage.
1
Sources
50%
Confidence
Long-term
Relevance
10/11/2026
First Seen
Sources
Related Entities
Related Claims
Unverified根据SWE-RPG基准论文，46.7%的代理失败发生在规划和需求分析阶段65% similarUnverifiedAccording to the properties of the binomial distribution, detecting a 5% failure rate with 95% confidence requires at least approximately 60 independent runs.65% similarUnverifiedTreating every failure as an equally weighted deduction is a serious evaluation pitfall, as an average score of 95% can mask the most costly and irreversible failures within that 5%.65% similarUnverifiedFor high-stakes fields such as finance and healthcare, 80% accuracy is not good enough.65% similarUnverifiedAmong the 91 failure cases, 51 were at the API layer and 25 were at the wrapper layer, accounting for the vast majority.64% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/998099API
curl https://kongchang.com/api/v1/knowledge/claims/998099MCP
get_claim(id=998099)