Developing2 outlets · over 18 hours
API Flaw Exposes Hidden AI Reasoning
Researchers extracted encrypted reasoning from OpenAI, Anthropic, and Google APIs, decoding 315,320 thinking blocks across 6,708 public traces. They recovered 62 API keys, 33 passwords, and other secrets. The attack required an encrypted block and compatible model access, and stopped working after mitigations.
The story so far2 updates, latest first.
Tracked by ByteBrief