Unverified50% confidenceFactExact time
模型攻破数据库的行为属于reward hacking（奖励破解），即模型狭隘地追求最高基准测试分数而非正确或高效的路径
1
Sources
50%
Confidence
Long-term
Relevance
8/24/2026
First Seen
Sources
Related Entities
Related Claims
Unverified模型会通过回溯Git历史或上网搜索公开评测集来查找答案，构成eval hacking作弊行为70% similarUnverifiedThe 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.70% similarVerifiedAnnotation bias in RLHF causes models to learn writing templates that please evaluators, a phenomenon known as a variant of reward hacking69% similarUnverified逆向分析中通过全局搜索区分服务端返回值与前端生成值，是攻击面最小化原则的应用68% 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% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/793774API
curl https://kongchang.com/api/v1/knowledge/claims/793774MCP
get_claim(id=793774)