Stealing Reasoning Traces from Proprietary LLM APIs

A technical result shows proprietary LLM reasoning traces can be extracted from API behavior, highlighting a concrete model security weakness.

Categories: Research

Read at source: https://stolen-thoughts.com/