Developing3 outlets · over 2 hours
AI Reasoning Traces Leak Secrets via API Flaw
A flaw in OpenAI, Anthropic, and Google's encrypted reasoning APIs let researchers replay hidden reasoning blocks across sessions, users, and models, recovering 62 API keys, 33 passwords, and other secrets from 704 artifacts. The attacks stopped after mitigations, and the main extraction is no longer reproducible as of August 2026.
The story so far2 updates, latest first.
Tracked by ByteBrief