Safety
An API flaw exposed big models' reasoning
🧒 Discovery
Specialists found a security hole in the big AIs and got it fixed. Phew!
📰 Essential
Researchers reveal a flaw at OpenAI, Anthropic and Google letting weaker models decode stronger models' reasoning. It has been fixed.
🔬 Expert
The mechanism: the APIs return hidden reasoning as encrypted blocks meant to be replayed from one call to the next, and those blocks could be replayed in another session, where a smaller model of the same family acted as a "fuzzy decoder". Across 6,708 public agent traces, the researchers recovered 62 API keys and 33 passwords. Practical advice: strip reasoning blocks from any shared trace.