OpenAI compaction flaw exposes hidden model reasoning
🛡️ Researchers uncovered a vulnerability in OpenAI's API compaction feature that allowed extraction of raw reasoning traces from frontier models. Xint published the technique, called compaction smuggling, on October 26 and reported OpenAI patched the issue on October 23 after confirming it and awarding a $6,000 bounty. The flaw enabled one model's hidden reasoning to be embedded in a compaction and later reconstructed by other models, undermining OpenAI's safeguards against distillation. Xint demonstrated the attack across multiple model versions and raised concerns about large-scale distillation risks.
