Unverified50% confidenceFactExact time
The historical attack memory component allows the system to retain recognition capability even when an attack pattern is inactive, enabling rapid invocation of historical experience to intercept the attack when the technique reappears
1
Sources
50%
Confidence
Long-term
Relevance
9/11/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedThe memory system includes self-healing mechanisms such as approximate deduplication, contradiction handling (retaining new facts and archiving old ones), decay mechanisms, and never hard-deleting data71% similarUnverified如果LoRA权重被逆向分析，攻击者可能从中推断出训练所用的私有数据内容，即记忆提取问题67% similarUnverifiedRAG通过以检索代替记忆，规避大模型知识过时问题并缓解模型幻觉67% similarVerifiedIndirect prompt injection conceals attack instructions within external documents, web pages, or database records retrieved by the model, and is activated when a RAG system introduces them into the context66% similarUnverifiedThe erroneous suspensions used a soft-ban mechanism, leaving users' conversation histories and custom settings data intact.64% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/898425API
curl https://kongchang.com/api/v1/knowledge/claims/898425MCP
get_claim(id=898425)