Unverified50% confidenceFactExact time
基于LLM的系统天然面临提示词注入（Prompt Injection）攻击
1
Sources
50%
Confidence
Long-term
Relevance
8/12/2026
First Seen
Sources
Related Claims
UnverifiedLLM-as-Attacker范式用一个专门微调的攻击LLM持续生成对抗提示，再由裁判LLM评估目标模型响应是否违规，形成无需人工介入的闭环78% similarUnverifiedBackdoor attacks plant triggers in training data, causing the model to behave normally on regular inputs but produce preset malicious outputs when specific trigger words are encountered.62% similarUnverifiedResearch from Carnegie Mellon University shows that even the most advanced models can be compromised by automatically generated adversarial inputs.62% similarUnverified某些越狱攻击的有效性来自于诱导模型进入特定吸引子状态，使输出在该框架内自我强化61% similarUnverifiedagentic LLM区别于普通LLM的本质特征是ReAct范式，即观察—思考—行动—观察的循环60% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/735463API
curl https://kongchang.com/api/v1/knowledge/claims/735463MCP
get_claim(id=735463)