Encrypted Chain-of-Thought of Top AI Models Cracked, Leaking 62 API Keys
xiaohu · x · 2026-08-13
A European research team (paper posted Aug 10) identified a scalable decryption jailbreak method to extract the concealed step-by-step reasoning traces from leading models by OpenAI, Anthropic, and Google.
Vulnerability Mechanism
The encryption algorithm itself isn't broken; rather, an architectural flaw exists in the API design: encrypted blocks are fully compatible and interchangeable across different sessions, users, and models within the same provider. Attackers can inject an encrypted reasoning trace from a target model into a weaker, less safeguarded model from the same provider, forcing it to decode and output the plaintext.
Practical Impacts
- Anti-distillation circumvention: Adversaries can easily extract a proprietary model's reasoning logic for piracy and distillation.
- Massive privacy leakage: Developers frequently share session logs publicly. By decoding 315,320 encrypted reasoning blocks scraped from public repositories, the team recovered 367 Personally Identifiable Information (PII) artifacts and 62 real API keys.
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