In a disclosure published October 26, 2026, security firm Xint described a way to pass reasoning from its own gpt-5.6-sol requests through OpenAI’s compaction feature so other models could echo the compacted text. Xint said OpenAI deployed a server-side fix on October 23, three days before the disclosure.
What compaction does
OpenAI describes compaction as a way to reduce context size while preserving state for later turns. The resulting item is encrypted and opaque, rather than intended for people to read. OpenAI’s API documentation explains how compaction works.
How Xint’s demonstration worked
Xint said its test used gpt-5.6-sol to generate reasoning and compact it, then passed the compacted item to gpt-4.1 and gpt-4o, which echoed its contents. The prompts were written by Xint researchers. The demonstration concerned their own requests; it did not demonstrate access to another user’s reasoning.
What the example establishes
In one example, Xint reported a 405-token reasoning item, a 612-token compaction item, and 609 tokens in each echoed output. Xint said checks with both echo models supported the repeatability of that step. The repeated text was the compacted item, however, and Xint said compaction may have paraphrased the original reasoning. The echoes therefore did not prove word-for-word recovery of the original.
Xint reported the fix date as October 23, 2026, and described its work as a demonstration involving its own prompts.
