Stealing reasoning traces from proprietary LLM APIs: the encryption tradeoffs nobody solved

bendee983 · x · 2026-08-19

A software engineer's take on the "Stealing Reasoning Traces from Proprietary LLM APIs" study: why did providers overlook basics like avoiding global encryption keys, and why send reasoning traces to clients at all?\n\nIn practice there are real tradeoffs: providers want to prevent distillation (hence encryption), reuse conversations across models (hence one key), and minimize client data stored server-side (hence shipping encrypted traces to clients). Trying to have it both ways led to a cascade of problems, including smaller, less safe models being exposed.

Original post →

More from Safety

Safety channel →