Encrypted Reasoning in Closed Models is 100% Recoverable, Paper Shows
Dany0 · reddit · 2026-08-11
A Reddit user shared a paper indicating that the encrypted reasoning processes of closed-source models like OpenAI's are 100% recoverable. The author urged the community to upload millions of traces from models like Anthropic's Claude to Hugging Face before providers patch the exploit.
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