Unverified50% confidenceFactExact time
Safety-aligned language models have a one-dimensional direction in activation space associated with refusing harmful requests.
1
Sources
50%
Confidence
Long-term
Relevance
10/23/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedCross-lingual jailbreaking exploits the asymmetry between a model's concept space and its language space, as the learning of safety-refusal behavior is highly language-specific.79% similarUnverifiedUnlike traditional software security vulnerabilities, the boundaries of language model jailbreaks are inherently fuzzy, and the same output may be judged as safe or dangerous depending on the context.76% similarUnverifiedLarge language models that have undergone safety alignment training often exhibit over-refusal behavior, erroneously refusing instructions that are themselves harmless but superficially touch on safety-related topics.75% similarUnverifiedDeterministic guardrails can only handle fixed-pattern content that looks like sensitive information and cannot detect natural language risks where sentences are fluent but subtly guide users toward incorrect actions.75% similarUnverified大语言模型在生成摘要时天然倾向于反映评论中的多数意见，可能将低频的安全警告边缘化72% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1088772API
curl https://kongchang.com/api/v1/knowledge/claims/1088772MCP
get_claim(id=1088772)