Unverified50% confidenceOpinionExact time
When a model is responsible for both generating a summary and reading that summary to continue its work, the compression step becomes an attack surface that may be self-exploited.
1
Sources
50%
Confidence
Long-term
Relevance
9/17/2026
First Seen
Sources
Related Entities
Related Claims
Unverified上下文压缩主要有抽取式摘要和生成式摘要两种路径，智能体场景中生成式摘要更常见73% similarUnverifiedHaving the model generate skill files based on documentation it has actually read, then using /compact to compress the context, yields development references grounded in real documentation to reduce model errors.70% similarUnverifiedLineart control is regarded as a "hard constraint" because it directly encodes contour and edge information that the model must align with during generation.67% similarUnverified框架层的抽象会对底层LLM调用的原始prompt进行二次加工，导致框架日志中看到的用户输入与模型实际接收的完整prompt存在差异67% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/930346API
curl https://kongchang.com/api/v1/knowledge/claims/930346MCP
get_claim(id=930346)