Researchers Extract AI Reasoning Traces via Weaker Sibling Models
Get the Tech newsletter
Daily tech — startups, AI labs, chips, the launches that shape the next decade. Free.
- Researchers devised a method to extract encrypted "reasoning traces" from frontier AI models including Claude, GPT, and Gemini by feeding them to a weaker model from the same provider.
- Frontier models' encrypted reasoning traces were successfully rendered into plaintext when processed through a weaker sibling model, exposing what the models internally deliberate.
Why it matters: If a weaker model can decode a frontier model's private reasoning, the safety and intellectual property value of encrypted traces across Claude, GPT, and Gemini is weaker than providers assumed — exposing a cross-provider vulnerability.
Ask SkimNews
