Researchers Extract AI Reasoning Traces from Claude, GPT, Gemini
Get the Tech newsletter
Daily tech — startups, AI labs, chips, the launches that shape the next decade. Free.
- Researchers devised a way to extract "reasoning traces" from Claude, GPT, and Gemini by feeding a frontier model's encrypted reasoning traces to a weaker model from the same provider, causing the weaker model to output the traces in plaintext.
- The finding demonstrates that encrypted reasoning features across three major AI providers' frontier models can be circumvented simply by routing the traces through the providers' own weaker models rather than breaking the encryption itself.
Why it matters: If encrypted reasoning traces are recoverable this easily — by passing them through the same provider's cheaper model — the reasoning features Claude, GPT, and Gemini have been marketing offer less protection of their internal chain-of-thought logic than the framing of 'encrypted' traces implies, potentially exposing proprietary model reasoning to anyone with API access to a lower-tier model.
Ask SkimNews
