Researchers Extract Encrypted Reasoning and Leaked Passwords from LLM APIs
The Decoder · rss · 2026-08-12
Security researchers uncovered a vulnerability in the APIs of OpenAI, Anthropic, and Google that allows the extraction of encrypted reasoning traces and their transfer across different models. A scan of public sessions revealed dozens of leaked passwords and API keys. Furthermore, the traces indicate that the reasoning summaries presented to users often hide what the models are actually doing.
More from Safety
- Vercel CTO: Open-Weight Models Lack Guardrails, Everything Hackable Will Be Hacked — cramforce · 2026-08-12
- What People Actually Worry About Regarding AI Data Centers — iamrobotbear · 2026-08-12
- Securing AI Agents: Avoiding Persistent Credential Leaks — davidcrawshaw · 2026-08-12
- Will AI Compute Runs Dictate Global Economy? A Deep Dive into Anti-Capitalism and AI Control — jessi_cata · 2026-08-12
- Apollo Red-Teams Anthropic's Auto Mode, Cutting Classifier Miss Rate to 7% — MariusHobbhahn · 2026-08-12
- New USENIX Paper Achieves True Messenger Privacy via Math, Hiding Recipients from Server — matthew_d_green · 2026-08-12