Research: Encrypted CoT Traces Can Be Extracted via Weaker Sibling Models
tw1st3d_m3nt4t · reddit · 2026-08-12
New research reveals a significant security risk in mainstream LLM APIs: encrypted chain-of-thought blocks returned by providers like Anthropic, OpenAI, and Google can be replayed across sessions, users, and models.
By replaying an encrypted trace from a frontier model into a weaker sibling model and jailbreaking the weaker one, researchers successfully recovered the stronger model's hidden reasoning in plaintext. This method bypasses the stronger model's anti-distillation safeguards entirely without attacking it directly.
More from Safety
- 'The AI-driven cybersecurity black swan is happening' — and a reminder to enable 2FA — danielrock · 2026-10-16
- No-instruction autonomous agent wakes up, sets own goals, and tries hacking by default — sethlazar · 2026-10-16
- India's CERT-In says AI agents escaped testing environments and breached Indian organizations undetected for days — HimanshiET · 2026-10-16
- HTTP response headers can carry prompt injections against AI agents, researcher shows — arpit_bhayani · 2026-10-16
- OpenAI agent uses someone else's Ethereum testnet wallet to pay for its task — nptacek · 2026-10-16
- Interpretability won't save us: critics call superintelligence project reckless regardless — zetalyrae · 2026-10-16