Unverified50% confidencePredictionExact time
After the model is optimized to be more proactive, attacks that were previously avoided through laziness will reappear, manifesting as security regressions.
1
Sources
50%
Confidence
Long-term
Relevance
9/11/2026
First Seen
Sources
Related Entities
Related Claims
Unverified推测执行过程中即使指令最终被回滚，其对缓存状态的影响仍会残留，攻击者可通过测量访存时延推断敏感数据72% similarUnverifiedReAct架构的核心循环为Thought→Action→Observation，攻击者可以在Observation阶段注入恶意内容72% similarUnverifiedThe model contamination caused by the NemoClaw attack is persistent—the model continues to exhibit malicious behavior after a normal restart, even after the initial attack vector has disappeared.70% similarUnverified大模型安全对齐本质上是在有用性与无害性之间寻找动态平衡，护栏过严损害用户体验，过松则留下安全漏洞70% similarUnverifiedCompared to the traditional plan-execute separation pattern, the greatest advantage of the ReAct paradigm lies in fault tolerance, allowing dynamic correction of subsequent plans based on tool return results.69% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/901694API
curl https://kongchang.com/api/v1/knowledge/claims/901694MCP
get_claim(id=901694)