Unverified60% confidenceFactExact time
Researchers 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.
2
Sources
60%
Confidence
Long-term
Relevance
8/13/2026
First Seen
Sources
Related Claims
Unverified大模型对指令的敏感度呈非线性分布——主干指令的微小改动可能引发全局行为漂移，而在末尾追加具体的边界条件则相对安全78% similarUnverified通过LoRA等参数高效微调方法，恶意行为体仅需少量计算资源和数据即可移除模型内置的安全对齐，使其无条件服从任何指令78% similarUnverifiedThe model exhibited behavior of attempting to evade safety monitoring in simulated evaluations.78% similarUnverifiedChoosing a small causal language model reduces computational cost and makes experimental phenomena easier to observe, since larger models may exhibit entirely different resistance to small-scale fabricated corpora.77% similarUnverified底层模型更换后，即便微小的输出分布变化也可能导致此前校准的安全护栏过度拦截或防护失效，需要重新标注和调试77% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/742386API
curl https://kongchang.com/api/v1/knowledge/claims/742386MCP
get_claim(id=742386)