Featured Post

OpenAI Cuts Reasoning Theft Attacks From 54 Percent to 0.2 Percent in GPT 6.2 Astra


OpenAI published a new safety post on 20 November 2026, and its headline number is striking. Attacks that try to steal the hidden reasoning inside GPT 6.2 Astra now succeed only 0.2 percent of the time, down from 54 percent against GPT 6 Astra. The company calls the trick adversarial distillation, and it says the same defenses now protect GPT 6.2 Sol as well.

The stakes are bigger than model theft. OpenAI says a stolen model can carry advanced cyber skills without the safeguards that govern the original, so a copy could help with attacks that the real model would refuse. That is why the company treats its chain of thought as its most valuable secret.

What happened

Adversarial distillation means using a frontier model without permission to train a second model with similar skills. The attacker feeds the model prompts, collects its answers and its hidden reasoning, then trains a copy that has weaker safeguards or none at all. OpenAI says the hidden reasoning matters most, because it holds steps that never appear in the final answer.

OpenAI tested its defenses with automated attackers that change tactics based on each response. In that adaptive black box test across six attacker models, stealing reasoning from an earlier request fell from 54 percent to 0.2 percent, and stealing it from the current request fell from 22.9 percent to 0.9 percent. These are OpenAI's own internal results, not independent tests.

Attack typeGPT 6 AstraGPT 6.2 Astra
Steal reasoning from an earlier request54 percent0.2 percent
Steal reasoning from the current request22.9 percent0.9 percent


The defenses work on two levels. At the request level, adversarial training teaches the model to refuse extraction attempts, classifiers scan messages and tool calls for leaks, and API controls tie reasoning to the conversation that produced it. If an earlier message is rewritten, the API drops the reasoning that followed it, and older model families cannot read GPT 6.2 reasoning at all.

At the campaign level, OpenAI investigates coordinated extraction attempts and bans the accounts behind them. Its weekly enforcement actions against distillation campaigns have ranged from thousands to tens of thousands of accounts, with the rate rising sharply this past summer. The company also shares indicators of compromise with other model providers.

Why it matters

Reasoning traces are the most useful training data an attacker can steal. In OpenAI's own study, a student model trained on exact reasoning traces generally beat one trained only on summaries in cyber and coding tests after supervised fine tuning. Protecting those traces can therefore limit how much cyber skill transfers to a copy, even when the task itself is allowed.

The post also names names. OpenAI says it briefed Congress in February on activity indicating that DeepSeek tried to distill United States frontier models, and in September it reported disrupting a campaign tied to people linked to Moonshot AI, the maker of Kimi. The company backs laws that would give the United States government more tools to punish foreign actors who extract model capabilities.

For builders, the message is that distillation defense is now part of the product. Selected safeguard improvements are also being applied to GPT 6 Astra, and OpenAI published a checklist so other developers can protect reasoning across the full interaction. For the newest OpenAI models behind these numbers, read our post on GPT Live 1.1.

What to watch

Watch whether other labs publish similar numbers. Anthropic is fighting systematic distillation attacks too, and OpenAI says it will share more soon about recent campaigns it disrupted. Independent replication of the 0.2 percent figure would turn a company claim into a benchmark.

Watch the policy fight as well. OpenAI supports the Chip Security Act, the AI OVERWATCH Act and the MATCH Act, which would tighten export controls and sanctions on actors who extract capabilities from American models. Distillation is now a national security debate, not only a safety blog post.

Read the full post

OpenAI has not posted an official YouTube video about this update. The full post is on the OpenAI blog: Combatting adversarial distillation.

Sources

OpenAI: Combatting adversarial distillation, 20 November 2026

The Decoder: OpenAI makes it much harder for attackers to extract how its latest AI models think, 20 November 2026

Follow Learn Daily AI on Instagram at @learningdailyai for one new AI story every day.

Comments