News

OpenAI Cut Hidden-Reasoning Theft on GPT-6.2 Astra From 54 Percent of Attacks to 0.2 Percent

OpenAI said on November 20 that adaptive attacks recovered GPT-6 Astra hidden reasoning 54 percent of the time, and 0.2 percent on GPT-6.2 Astra with new safeguards.

OpenAI just published the numbers on a theft it has been chasing all year: stealing the hidden chain of thought from a frontier model so you can train a copy that does not have to obey the original rules.

The company said on November 20 that adaptive black-box attacks recovered protected reasoning from an earlier prompt 54.0 percent of the time against GPT-6 Astra with its original safeguards. Against GPT-6.2 Astra with the new stack, that rate fell to 0.2 percent. Pulling the reasoning from the current request fell from 22.9 percent to 0.9 percent. OpenAI is now applying some of those defenses back onto GPT-6 Astra.

The target is not the final answer. Protected reasoning is the intermediate work the model does before it replies. Those traces can carry steps the visible answer leaves out, which makes them useful training data for a second model. OpenAI's worry is a student that learns cyber or biology skills from a teacher that would have refused to use them that way.

Exact traces beat summaries

In an internal cyber study, OpenAI trained student models on Astra-level teachers doing coding and defensive security work. One student got the usual user-visible material: final replies, reasoning summaries, and tool calls. The other got the teacher's exact reasoning traces. After supervised fine-tuning, the exact-trace student generally scored higher. After a light round of reinforcement learning, it still led at the largest inference budget OpenAI tested.

The company treats that as a reason to keep the raw chain of thought off the wire. Defenses now sit at two layers. On each request, adversarial training, input and output classifiers, and API rules try to stop a leak, including in tool-call arguments. Across accounts, OpenAI looks for coordinated extraction campaigns and says weekly enforcement has run from thousands to tens of thousands of accounts, with the pace up this past summer.

GPT-6.2 Astra and GPT-6.2 Sol now tie protected reasoning to the conversation that produced it. Rewrite an earlier message and the API drops the reasoning that followed. Older model families with weaker safeguards cannot read Astra and Sol's traces. OpenAI already described GPT-6.2 Astra as a model it holds at a self-improvement threshold. The distillation post is the same lab arguing the hidden work has to stay hidden.

OpenAI names names, and asks Congress for more tools

OpenAI says it briefed Congress in February on activity indicating DeepSeek was trying to distill U.S. frontier models. In September it said it disrupted a campaign to extract protected reasoning and attributed a core part of that activity to people associated with Moonshot AI, the developer of Kimi. It says it plans to share more on recent campaigns.

The company wants legislation that would let the U.S. government identify and sanction foreign actors who extract capabilities from American models. It told Congress it supports the Chip Security Act, the AI OVERWATCH Act, and the MATCH Act. Export controls, in that telling, limit the chips. Distillation defenses limit a shortcut around the chips.

What the post does not show is an independent test of those attack-success rates, or a public sample of the campaigns. The 54.0 percent and 0.2 percent figures are OpenAI's own adaptive evaluation, aggregated across six attacker models. The claim is that the new stack makes the hidden reasoning much harder to steal. The copy, if someone still gets it, is the part that would not have to ask permission.

Some offers on this page may be paid placements or contain affiliate links.

Subscribe to Techpresso

Free daily newsletter, read in 5 minutes.

Subscribe free