Paper Reveals Encrypted CoT Flaw Allowing Reasoning Theft from Top LLMs
Alexander Panfilov · hf · 2026-08-11
A new research paper reveals an architectural vulnerability in the encrypted chain-of-thought (CoT) blocks used by leading LLM APIs (e.g., OpenAI, Anthropic, Google) to protect proprietary reasoning. These blocks are fully compatible and interchangeable across different sessions, users, and models within a provider's ecosystem.
Exploiting this, researchers developed a scalable decryption jailbreak: by injecting an encrypted reasoning trace from a stronger model into a weaker, less safeguarded model within the same ecosystem, the weaker model can be forced to decode and output the trace verbatim in plaintext. This enables four major attack vectors:
- Anti-distillation circumvention: Allows adversaries to extract a proprietary model's hidden reasoning.
- Large-scale PII extraction: By decoding 315,320 reasoning blocks scraped from public developer logs, researchers recovered 367 PII artifacts and 182 credentials.
- Hazardous information exposure: Reveals dangerous hidden reasoning even if the model's final output safely rejects the malicious request.
- Invisible prompt injections: Attackers can embed malicious payloads entirely within encrypted blocks to poison public agentic rollouts.
Following responsible disclosure, the authors proposed concrete cryptographic and system-level mitigations.
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