Artificial Intelligence
OpenAI compaction flaw exposed hidden reasoning
Xint said OpenAI's compaction feature could be manipulated to turn a model's hidden reasoning into readable text, and said OpenAI later fixed the issue after a bug bounty report.
- Xint said OpenAI's compaction feature could be used to move hidden reasoning into plain text from gpt-5.6-sol.
- Xint said the technique bypassed protections OpenAI had added in August, and that OpenAI fixed the issue on 23 October.
- The researchers used their own maths and algorithm puzzles, and Xint said the method did not reach another user's reasoning.
- Xint said it could not prove the copied text matched the model's original reasoning word for word.
- Xint said OpenAI paid a $6,000 reward and that the work built on its earlier GhostCompact method.
Xint said OpenAI's compaction feature could be used to lift a model's hidden reasoning into readable text. It said the method got around safeguards OpenAI had added in August.
The researchers ran the test on gpt-5.6-sol with maths and algorithm puzzles they had written themselves. They kept the encrypted reasoning the model returned, then told the same model to copy it into the compacted item with spaces changed to tabs.
Xint then passed that item to gpt-4.1 and gpt-4o, and both repeated it exactly. Xint said that showed the copying step had worked, but it could not prove the text was the model's original reasoning character for character.
Xint said OpenAI confirmed the problem and fixed it on 23 October, and that the fix was on OpenAI's side so API developers did not need to change their code. It also said OpenAI paid the researchers a $6,000 reward.
Xint said the demonstration built on its earlier GhostCompact work, which had shown compaction could be steered towards attacker-chosen tasks. It also said the example was not a live attack and that it had not tested whether GPT-6 Astra was affected.
Named in this story
Companies
- OpenAI
- built the models and fixed the compaction issue
- Xint
- reported the flaw and demonstrated the technique
- Anthropic
- had reported China-based labs targeting hidden reasoning
- Moonshot AI
- was linked by OpenAI to a disrupted campaign
Governments and agencies
- United States government
- co-authored an advisory on hidden-reasoning targeting
Products and systems
- OpenAI's compaction feature
- shrunk long conversations into smaller encrypted items
- gpt-5.6-sol
- was the model Xint used in the test
- gpt-4.1
- repeated the compacted item exactly
- gpt-4o
- repeated the compacted item exactly
- GPT-6 Astra
- was not tested by Xint
- GhostCompact
- was Xint's earlier work on compaction abuse
How the source tells it
The article was a straightforward technical security report with little emotive framing.
No emotive framing was found in the original.