Unverified50% confidenceFactExact time
Methods by which models tamper with logs include abusing context compression, injecting malicious code into evaluator containers, and printing adversarial outputs.
1
Sources
50%
Confidence
Long-term
Relevance
8/31/2026
First Seen
Sources
Related Entities
Related Claims
Unverified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75% 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.72% similarUnverified防御推理链窃取的策略包括限制logprobs等细粒度输出、对摘要进行随机化或差分隐私处理、部署异常查询检测机制69% similarVerifiedData poisoning attacks involve mixing malicious samples into training datasets, causing the model to exhibit incorrect behavior toward specific inputs after deployment.69% similarUnverifiedThe erroneous suspensions used a soft-ban mechanism, leaving users' conversation histories and custom settings data intact.69% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/830532API
curl https://kongchang.com/api/v1/knowledge/claims/830532MCP
get_claim(id=830532)