Study Confirms: API Vulnerabilities Expose Hidden CoT in Frontier Models, Enabling Cross-Model Transfer
gsarti_ · x · 2026-08-13
Researchers have discovered a vulnerability in the APIs of every frontier AI company that allows the extraction of hidden chain-of-thought (CoT) reasoning. They verified that the extracted reasoning token count matches the billed API thinking tokens 1:1 for most queries.
Furthermore, independent research provides evidence for cross-model CoT transfer: transferring CoT from a stronger model to weaker models makes the weaker models' performance closely match the strong one. This generalization correlates with human preference rankings and RL post-training, prescribing caution when using LRM explanations for new insights.
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