NewsIncident
Compaction smuggling: a three-step trick pulled hidden reasoning out of OpenAI's models for a 6,000 dollar bounty
Xint, part of Theori, published on October 26 a technique it calls compaction smuggling that leaked the raw reasoning traces of OpenAI's frontier models. A task goes to gpt-5.6-sol with store set to false and the encrypted reasoning item is kept; the task is removed from the history and the /responses/compact endpoint is told to preserve the reasoning verbatim under a "solution" key, with every space replaced by a tab to slip past the usage-policy classifiers; then GPT-4.1 and GPT-4o are asked to output the compaction item unchanged. In the example run a 405-token reasoning item came back as two 609-token echoes, at every reasoning effort level from minimal to max. OpenAI's August mitigations after the Stolen Thoughts paper block replaying reasoning into a different model, but compaction runs on the same model, so they did not apply. Xint found the bug in August, OpenAI deployed an undisclosed fix on October 23 and paid 6,000 dollars. Xint places it in the distillation context of the autumn: the NSA, CISA and FBI advisory on industrial-scale distillation campaigns in early September, Anthropic's report of seven campaigns by China-based labs days later, and OpenAI's September 30 disclosure of a reasoning extraction campaign that peaked at about 16,000 requests from more than 4,000 accounts in July.
Why it matters
Every server-side step that reads a model's hidden reasoning and takes instructions from the client is a possible way to get that reasoning out. If you build on a provider's API, the same lesson applies to your own summarization, memory and compaction features.