Researchers identify vulnerability in OpenAI model reasoning traces

OpenAI recently fixed a security weakness in its application programming interface (API) that allowed people to extract raw reasoning traces from its most advanced AI models, according to Infosecurity Magazine.

The security weakness was found by a company called Xint. The researchers used a method called "compaction smuggling" to get around protections designed to keep the AI's hidden reasoning, also known as its chain of thought, secret. OpenAI usually gives API users summaries instead of raw reasoning to make it harder for others to copy how its models work. However, Xint found that the feature that compresses conversation history could be misused. By hiding reasoning inside a compressed conversation, the researchers got other models to reveal this hidden information.

In August, OpenAI added protections to stop people from replaying encrypted reasoning. However, these protections did not cover the process used to compress conversations. Xint showed that the attack worked against the GPT-5.6 Sol, GPT-4.1, and GPT-4o models. OpenAI confirmed the problem, released a fix on October 23, and awarded the researchers $6,000 through its bug bounty program. The discovery highlights continuing concerns across the AI industry about model distillation, which involves copying a model's capabilities, and the safety of its hidden reasoning.

Earlier reports from Anthropic and OpenAI have described similar attempts by outside groups to extract hidden information. These reports show why strong security protections are needed as advanced AI models become more complex and are used more widely across different industries.