Unverified50% confidenceFactExact time
2023年卡内基梅隆大学Andy Zou等人发表论文《Universal and Transferable Adversarial Attacks on Aligned Language Models》
1
Sources
50%
Confidence
Long-term
Relevance
7/12/2026
First Seen
Sources
The Boundaries of AI Safety Guardrails: Deep Reflections Sparked by a Single Prompt
redditr/ChatGPT7/11/2026
Related Claims
Unverified强对抗性攻击如使用另一个模型完整改写或翻译往返仍可能破坏水印信号，鲁棒性与文本质量之间存在权衡68% similarUnverifiedCross-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.68% similarUnverifiedOpenAI、DeepMind等机构的多项研究都证实了大语言模型谄媚问题的普遍性65% similarVerifiedThe ReAct pattern was formally proposed by Yao et al. in the 2022 paper "ReAct: Synergizing Reasoning and Acting in Language Models".65% similarUnverifiedIn multi-Agent collaborative development, large language models have a tendency to be "overly enthusiastic" and may overstep their responsibilities by modifying code outside their designated scope.64% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/490674API
curl https://kongchang.com/api/v1/knowledge/claims/490674MCP
get_claim(id=490674)