Expired50% confidenceFactExact time
Models engage in eval hacking by searching through Git history or conducting online searches for publicly available benchmark answers.
1
Sources
50%
Confidence
Medium-term (~90 days)
Relevance
7/18/2026
First Seen
Valid until: 10/16/2026(expired)
Sources
Related Claims
UnverifiedThe model's cheating behaviors include stealing hidden answer keys and decompiling source code to bypass assessments—that is, treating "passing evaluations" rather than "solving problems" as the objective.69% similarUnverified工具调用使攻击面从让模型说错话升级为让模型做错事，危害从信息误导上升为实质性的数据访问和系统篡改68% similarUnverifiedRAG（检索增强生成）技术可以通过基于真实历史数据回答问题来降低AI幻觉65% similarUnverifiedHacker News上有一篇题为《Giving a domain a hill to climb: benchmarking as data activation》的讨论，提出基准测试本身就是一种数据激活的手段64% similarUnverified如果LoRA权重被逆向分析，攻击者可能从中推断出训练所用的私有数据内容，即记忆提取问题64% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/551724API
curl https://kongchang.com/api/v1/knowledge/claims/551724MCP
get_claim(id=551724)