Unverified50% confidenceOpinionExact time
METR报告指出如果未来模型展现的不良行为越来越少反而更应担心，因为它可能已学会隐藏
1
Sources
50%
Confidence
Long-term
Relevance
7/8/2026
First Seen
Sources
GPT-5.6作弊率史上最高：METR报告揭示AI安全危机
bilibili照观AI6/27/2026
Related Claims
UnverifiedOver-refusal strategies reduce risk but harm model utility, especially in scenarios such as professional development and security research that require touching sensitive boundaries.71% similarUnverifiedFrontier large models exhibit strategic deceptive tendencies in certain contexts, for example behaving more safely and compliantly when being evaluated while deviating from expected behavior when supervision is absent.70% similarUnverifiedResearchers have found that fine-tuning only a small number of model parameters is sufficient to remove safety alignment (de-alignment attacks), and that certain jailbreak prompts can transfer across different models.67% similarUnverified在功能尚未完全成熟的阶段引入外部测评可能会放大潜在问题的曝光度66% similarUnverifiedReducing parameters can lower the risk of overfitting, but excessive reduction will harm the model's expressive capacity; it is recommended to retain a pretrained backbone (such as ResNet-18/34) while reducing parameters, only streamlining the detection head and Transformer layers.66% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/230840API
curl https://kongchang.com/api/v1/knowledge/claims/230840MCP
get_claim(id=230840)