Unverified60% confidenceEventExact time
Researchers Ilya and Sasha published a paper titled "Stealing Reasoning Traces from Proprietary LLM APIs," studying the theft of reasoning traces from proprietary large model APIs.
2
Sources
60%
Confidence
Long-term
Relevance
9/2/2026
First Seen
Sources
Related Entities
Related Claims
Unverified《Stealing Reasoning Traces from Proprietary LLM APIs》研究指出，攻击者可通过API返回的间接信号重构甚至窃取模型的完整推理轨迹80% similarUnverifiedAbsence verification requires the model to actively infer which important information that should have appeared in the source material was omitted, necessitating the prior construction of a complete mental model of what ought to be present, which is far more difficult for LLMs.62% similarUnverifiedThis test bypassed the retrieval step and fed all candidate documents directly to the models, thus revealing the reasoning capability bottleneck of language models themselves rather than defects in retrieval algorithms.61% similarUnverifiedEarly large models in cross-source data research tasks were prone to missing key sources, numerical confusion, or broken inference chains, forcing agents to repeatedly supplement queries.61% similarUnverifiedIn its official article on Contextual Retrieval, Anthropic gives the example that an embedding model may find generic content about an error code yet miss an exact "TS-999" match.61% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/843809API
curl https://kongchang.com/api/v1/knowledge/claims/843809MCP
get_claim(id=843809)