Researchers extract hidden reasoning traces from Claude, GPT, and Gemini
The extraction method works on all major API-accessed frontier models, per researchers at University of Tübingen, Max Planck, and Snyk. Kimi K3's reasoning closely matched Claude Opus 4.8 and GPT 5.6 Sol, suggesting possible distillation — though not causally proven. The flaw also allowed recovery of passwords and API keys, since fixed.
Featured · Alexander Panfilov
1 source
Daily brief
Get tomorrow's AI brief in your inbox
More stories today
- When AI Models Become Commodities — What Open-Weight AI Could Mean For LegalTech
- Semantic caching tool reduces LLM costs up to 80%
- Architecture-aware harness keeps AI-generated code on-structure
- River AI raises $1.1 billion to build personal AI platform
- Build an n8n AI agent to monitor your homelab with human approval