116-Page Paper Reveals Vulnerability: Extracting Encrypted Reasoning from Top LLMs
xeophon · x · 2026-08-11
A new 116-page paper highlights a significant vulnerability in models from OpenAI, Anthropic, and Gemini, allowing attackers to extract encrypted raw reasoning at scale.
This vulnerability leads to multiple security issues, including distillation attacks and credential extraction. The researchers also found numerous instances of illegible reasoning (especially in GPT models), unfaithful reasoning, and evidence suggesting certain open-weight models were likely distilled.
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