OpenAI says it has hardened GPT-6.2 Astra and GPT-6.2 Sol against adversarial distillation, the practice of copying a frontier model’s abilities through its outputs. In its own red-team tests, attacks that stole hidden reasoning fell from a 54.0% success rate to 0.2%.
What Adversarial Distillation Is
Distillation itself is a standard machine-learning technique: one model learns from another model’s answers, reasoning, or judgments. It turns hostile when done without permission, because the copy can inherit frontier abilities while shedding the safeguards that governed the original model.
OpenAI focuses on the model’s chain of thought, the intermediate steps it generates while working toward an answer. Those traces can hold information the final answer leaves out, which makes them especially valuable training data for anyone trying to reproduce the model’s skills.
In an internal study, OpenAI trained two student models from the same starting point, one on exact reasoning traces and one on summaries. The exact-trace student generally learned more cyber capability, showing that protecting traces can limit how much dangerous skill transfers to a copy.
How the Defenses Work
The defenses work at two levels. At the request level, adversarial training teaches the model to refuse extraction attempts while still explaining itself normally. Input and output classifiers watch for leaks, including in tool-call arguments, where reasoning can escape unseen.
At the API level, protected reasoning is now tied to the conversation that produced it. If an earlier message is rewritten, the reasoning that followed is discarded. Older model families with weaker safeguards can no longer read GPT-6.2 Astra and Sol’s protected reasoning at all.
At the campaign level, OpenAI hunts for coordinated extraction across accounts, sessions, and infrastructure, then enforces against whole networks. Weekly enforcement actions this year have ranged from thousands to tens of thousands of accounts, with the rate rising sharply over the summer.
Our Read
The headline number is striking, but it is OpenAI’s own test, not an independent verdict. A 0.2% success rate in red-teaming says the wall holds against the attacks OpenAI built; it does not prove no attacker will find a new route. The honest read is narrower: extraction got more expensive.
That still matters, because the fight is now about economics as much as engineering. OpenAI names DeepSeek and Moonshot AI in the same breath as Congress briefings and new legislation, turning a technical fix into a national-security argument. Rivals must now spend more to copy less.
OpenAI says it shares indicators of compromise with other model providers so they can hunt the same abuse, and it works through the Frontier Model Forum. The industry is quietly building a shared wall, even as its members compete fiercely on everything else in public.
Why It Matters
If the defenses hold, frontier labs keep the value of their most expensive asset: the reasoning inside their models. If they fail, capabilities built under strict safety rules can spread to models with none. For users, the stakes are quieter — safer models, and fewer copied ones.
Background
The publication follows OpenAI’s September 30 disclosure of a coordinated campaign observed from July 1. That campaign peaked with 16,000 extraction requests from more than 4,000 users across July 24 and 25, and OpenAI attributed a core cluster to individuals associated with Moonshot AI.
Source: OpenAI’s post on combatting adversarial distillation


0 Comments