OpenAI GPT-6.2 Locks Down Its Reasoning to Stop Rivals Copying It

A person types a question into ChatGPT on a laptop

OpenAI has a copying problem — and it is not the kind that ends in a lawsuit. Rival labs, the company says, have been quietly training cheaper copies on its models’ outputs. Its new account of the fight carries a startling number: attacks that once stole a model’s hidden reasoning 54% of the time now succeed 0.2% of the time.

The defenses, described in a Nov. 20 post on OpenAI’s own site, ship in GPT-6.2 Astra and GPT-6.2 Sol. They target what OpenAI calls adversarial distillation: training a copy on a frontier model’s outputs, without authorization, minus its safeguards.

The Heist OpenAI Is Trying to Stop

Distillation itself is ordinary machine learning. It turns hostile when the student inherits the teacher’s capabilities but not its safety rules: a model that helps defenders find software flaws could, distilled into an unguarded copy, help attackers exploit them — and one that assists with biological research could help build a harmful pathogen.

The tests focus on the most valuable thing rivals can steal: the model’s protected reasoning, the chain-of-thought steps it generates while working toward an answer. In OpenAI’s adaptive black-box evaluation — automated red-teamers attacking as an outside user would, across six attacker models — the success rate for recovering reasoning from an earlier request fell from 54.0% on GPT-6 Astra to 0.2% on GPT-6.2 Astra. For the current request, it fell from 22.9% to 0.9%.

How the Lock Works

The core trick is to tie reasoning to the conversation that produced it. In GPT-6.2 Astra and Sol, protected reasoning stays attached to its chat history: rewrite an earlier message and the API drops the reasoning that followed. Older model families cannot read the new models’ reasoning.

Underneath sit two more layers: adversarial training that teaches the model to refuse extraction attempts while still explaining things to legitimate users, and classifiers screening for attempts and leaks — including leaks hidden in tool-call arguments. OpenAI also investigates coordinated campaigns, banning thousands to tens of thousands of accounts weekly, a rate that rose sharply over the summer.

Why This Matters

Here is the finding the announcement buries beneath its own good news. OpenAI studied how much cyber capability transfers when a student trains on a teacher’s outputs: the exact reasoning traces matter more than the answers. A student trained on exact traces generally outperformed one trained on summaries, even on benign tasks. The chain of thought is the capability, not just a record of it — the difference between a model that helps find vulnerabilities and one that helps exploit them.

The numbers carry a caveat most coverage skipped: they are OpenAI’s own internal tests, run by its own red-teamers, with the six attacker models unnamed and no independent verification. The 54-to-0.2% drop measures OpenAI’s defenses against OpenAI’s attacks. Whether it holds against a determined rival, the industry must take on trust — while Anthropic fights the same kind of attacks, The Decoder reports.

A Technical Fix With a Political Tail

OpenAI is not presenting this as engineering alone. The same document names the rivals it believes are doing the copying: in February it briefed Congress on activity indicating DeepSeek was attempting to distill U.S. frontier models, and in September it said it disrupted a campaign and attributed a core part of it to people associated with Moonshot AI (Kimi). It also backs three bills — the Chip Security Act, the AI OVERWATCH Act, and the MATCH Act — to give Washington more tools against foreign extraction.

The pairing is deliberate: export controls limit rivals’ access to training hardware, and distillation defenses make it harder to acquire the same capabilities from models already trained. GPT-6.2 Astra and Sol launched on Nov. 17, with Astra’s API prices cut 20% and both models rated at the Critical level for cyber capability, BetaNews reports. OpenAI is shipping its most capable models, and its most locked-down ones, in the same release. The message to rivals: the reasoning is no longer free. The message to Washington: the locks need laws behind them.

Meanwhile in AI Tech: Anthropic built its own anti-distillation safeguards for Claude — and OpenAI scrapped GPT-6.1 Astra over safety concerns.

Written by
Ryan covers artificial intelligence and enterprise tech — from foundation models and AI chips to the business of machine intelligence. He tracks model releases, funding rounds, and the policy moves shaping the AI industry.