Unverified50% confidenceOpinionExact time
The multi-layered safety guardrails of modern LLMs are not a unified logical system but rather multiple independent tendencies formed during training, and different safety rules triggered by different contexts may produce contradictory behavior.
1
Sources
50%
Confidence
Long-term
Relevance
8/9/2026
First Seen
Sources
AI幻觉实录：Perplexity假装登录了你的账户
Related Claims
Unverified当前对齐方法主要依赖训练阶段的行为塑造，在面对精心设计的对抗性提示时安全边界可能被突破，这种现象称为'越狱'（jailbreaking）71% similarUnverified通用越狱指单一提示词或操作序列能系统性绕过模型所有安全约束，而针对性越狱只能解锁某一类具体有害行为67% similarUnverified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66% similarUnverified越狱技术通过特定提示工程手段绕过模型内置安全护栏，使其输出本应被拒绝的内容66% similarUnverifiedBehavior Cloning (BC) suffers from the covariate shift problem: during training the model only sees the state distribution of expert trajectories, and at deployment small errors can drive the robot into states not covered by the training set, leading to cascading failures.65% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/717058API
curl https://kongchang.com/api/v1/knowledge/claims/717058MCP
get_claim(id=717058)