Unverified50% confidenceFactExact time
Forms of information leakage in code evaluation include test case visibility leakage, environment metadata leakage, and training data contamination.
1
Sources
50%
Confidence
Long-term
Relevance
8/6/2026
First Seen
Sources
AI Agent奖励黑客：为什么评估高分不等于真实能力
Related Claims
Unverified安全检测层往往在代码进入沙箱执行之前就基于代码外表而非实际危害进行分析，导致误报70% similarVerifiedTraining data contamination is the primary challenge faced by AI programming benchmarks, as test cases come from public GitHub repositories, which are the same sources used as pretraining data for large models.69% similarUnverifiedPublicly available safety evaluation sets carry the risk of being included in training data and causing evaluation contamination, which includes two risks: model vendors gaming scores and malicious actors studying test sets to develop bypass variants.69% similarUnverifiedCommon methods for detecting data contamination include verbatim completion tests and membership inference attacks.67% similarUnverifiedCode references in an SDK can be one of three situations: a genuine product preview, internal placeholder/test code, or distorted information dissemination.67% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/698548API
curl https://kongchang.com/api/v1/knowledge/claims/698548MCP
get_claim(id=698548)