Unverified50% confidenceOpinionExact time
This research extends the LLM safety discussion from "what the model says" to "how the model says it," pointing out that every piece of model-generated text users share can become a vehicle for information leakage.
1
Sources
50%
Confidence
Long-term
Relevance
11/13/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedIn LLM agent scenarios, the model treats tool description fields as part of its own configuration to follow and treats result fields as external data to process, and this difference in trust level causes the same text to carry different injection risk in different fields.72% similarUnverifiedModel behavior is shaped jointly by training data, prompts, context, and external tool calls, and contamination or misuse of any one link can lead to harmful output, leakage of sensitive information, or unauthorized operations.69% similarUnverifiedSystem prompt leakage stems from how the model works — the model reads and repeats back the text in front of it, rather than being some code defect that can be patched.69% similarUnverifiedWhen an LLM generates an answer, it involves information sources at multiple levels: knowledge already encoded in the model weights, web content retrieved in real time, and trustworthy information filtered out through citation and consensus mechanisms.69% similarUnverifiedThe model absorbs the content set in the system prompt as narrative truths about the character, so dialogue examples need to be filled with factual statements.68% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1166005API
curl https://kongchang.com/api/v1/knowledge/claims/1166005MCP
get_claim(id=1166005)