Stolen Thoughts: Two API Calls Extract Frontier Model Reasoning; Anthropic Rolls Out Preserved Thinking
zainhas · x · 2026-10-13
The 'Stolen Thoughts' paper shows Anthropic, OpenAI and Google return encrypted chain-of-thought blocks that can be replayed across sessions, users and models. By replaying a frontier model's trace into a weaker sibling and jailbreaking it, researchers recover the stronger model's hidden reasoning in plaintext in just two API calls—without touching the strong model or tripping anti-distillation safeguards. A follow-up audit found verbatim extraction still possible via third-party API providers. Anthropic's response: 'preserved thinking' on Claude Fable 5.1+—the API checks a thinking block's signature so only the producing model or a fixed whitelist (Fable 5.1 reads Opus 5/5.5, not vice versa) can use it, silently dropping blocks otherwise.
More from Models
- Meta's Muse Wonder (Max) tested on one-shot 3D scenes and front-end vs top models — arena · 2026-10-13
- Five Ways to Monetize Open-Weight LLMs, API Sales Lead — Terminator857 · 2026-10-13
- Quintin Pope: pretraining is a degenerate special case of RL, subsumed by RL with right data — QuintinPope5 · 2026-10-13
- Muse Code coding model impresses in early testing: faster than Claude Code — QuixiAI · 2026-10-13
- 1.2M Calls Study: Even at Temperature 0, LLM Benchmarks Vary Up to 24% — lulzxdxdxd · 2026-10-13
- "Another benchmaxxed release in 14 hours": AI circle mocks benchmark hype — thesaraharminta · 2026-10-13