News

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

OpenAI said November 20 that adaptive attacks recovering GPT-6.2 Astra hidden reasoning fell from 54.0 percent on GPT-6 Astra to 0.2 percent. Current-request theft fell to 0.9 percent.

OpenAI spent months chasing people who train knockoff models on ChatGPT's hidden scratch work. On November 20 it published the score: those thefts used to work more than half the time.

On November 20, OpenAI said it had rebuilt the defenses that hide GPT-6.2 Astra's and GPT-6.2 Sol's protected reasoning, the internal chain-of-thought the final answer does not show. In an internal adaptive test, automated red-teamers tried to recover reasoning the model had generated on an earlier prompt. Average attack success fell from 54.0 percent against GPT-6 Astra with its original safeguards to 0.2 percent against GPT-6.2 Astra with the new stack. Stealing reasoning from the current request fell from 22.9 percent to 0.9 percent. Both figures are black-box tests, aggregated across six attacker models.

That is the same GPT-6.2 Astra Techpresso covered when OpenAI called it the first externally released model to meet its High threshold for AI self-improvement.

Why the scratch pad is the prize

Distillation is how one model learns from another: answers, judgments, or the reasoning in between. OpenAI's worry is not only a copied product. A student trained on a teacher's cyber or biology traces can pick up skills the original model is not supposed to hand to an attacker. The company ran that experiment on coding and defensive cybersecurity tasks. After supervised fine-tuning, the student that saw exact reasoning traces generally beat the student that saw summaries. After a light round of reinforcement learning, the exact-trace student still scored higher at the largest inference budget they tested.

The Moonshot-linked extraction Cyberpresso covered is the live version of that lab. OpenAI says it disrupted a bid to extract protected reasoning and attributed a core part of it to people tied to Moonshot AI, the Kimi developer. In February it briefed Congress on DeepSeek attempting to distill US frontier models.

Request, then campaign

OpenAI now trains the model to refuse extraction attempts, runs classifiers on inputs and on outputs including tool-call arguments, and ties GPT-6.2 Astra and Sol reasoning to the conversation that produced it. Rewrite an earlier message and the API drops the reasoning that followed. Older model families cannot read the new traces. Weekly enforcement against distillation campaigns has ranged from thousands to tens of thousands of accounts, with the rate rising over the summer.

The national-security pitch sits next to export controls: stop a foreign actor from reproducing US model capabilities without the refusals. OpenAI told Congress it supports the Chip Security Act, the AI OVERWATCH Act, and the MATCH Act. It also says it is applying some of the new request-level fixes back to GPT-6 Astra.

The 0.2 percent is an internal red-team number, not a claim that distillation is over. OpenAI still sees campaigns. The change is that the hidden reasoning is harder to walk out the door with.

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