OpenAI API Flaw Exposes AI Model Reasoning Traces

Researchers discovered a vulnerability in the OpenAI API, called "compaction smuggling," that let them extract raw reasoning traces from AI models, information OpenAI normally keeps hidden. OpenAI has fixed the flaw and paid the researchers a $6,000 bug bounty.

Researchers uncovered a vulnerability in the OpenAI API, dubbed "compaction smuggling," which allowed them to extract raw reasoning traces from AI models - information that OpenAI typically keeps under wraps. The flaw has since been fixed, with OpenAI paying a $6,000 bug bounty to the researchers who discovered it.