OpenAI has put a number on its fight against model theft — and named the rivals it says are doing the stealing. In a post published on Friday, the company said attacks designed to pull a model’s hidden reasoning out of GPT-6.2 Astra now succeed 0.2% of the time, down from 54% against the previous generation, and it publicly tied extraction campaigns to DeepSeek and to people associated with Moonshot AI.

The post, titled “Combatting adversarial distillation,” is the most detailed account OpenAI has given of how it defends its models — and the most openly geopolitical. Distillation itself is an ordinary machine-learning technique: one model learns from another’s outputs. It becomes “adversarial,” in OpenAI’s framing, when those outputs are taken without permission to build a rival that inherits the capability but not the safeguards.

What OpenAI measured#

The headline figures come from OpenAI’s own testing. In an adaptive black-box evaluation — automated attackers that adjust their approach as the system responds, seeing only what an outside user would see — attempts to recover protected reasoning from an earlier request succeeded 54.0% of the time against GPT-6 Astra with its original safeguards, and 0.2% against GPT-6.2 Astra with the updated ones. For reasoning from the current request, the rate fell from 22.9% to 0.9%. The results are aggregated across six attacker models.

Attack typeGPT-6 Astra, original safeguardsGPT-6.2 Astra, updated safeguards
Extracting protected reasoning from an earlier request54.0%0.2%
Extracting protected reasoning from the current request22.9%0.9%

Three caveats matter. The numbers are OpenAI’s own, from its own test harness; no independent verification has been reported. The six attacker models are not named. And there is no per-domain breakdown, so the figures say nothing about which kinds of reasoning — code, science, language — are hardest to extract. OpenAI says it is now applying selected safeguard improvements to GPT-6 Astra as well.

Why the hidden reasoning is the prize#

The thing OpenAI is guarding is not its answers. It is the working-out: the chain-of-thought steps a model generates internally before it answers, which users see only in summary form. Those traces can contain information the final answer leaves out, which makes them especially valuable as training data for a second model.

OpenAI’s own study on cyber capability transfer makes the stakes concrete. It trained two student models from the same starting point — one on the teacher’s exact reasoning traces, the other on the summaries users normally see. After supervised fine-tuning, the exact-trace student generally performed better; after further reinforcement learning, it still scored higher at the largest inference budget tested on both benchmarks. The conclusion OpenAI draws: protecting exact traces can limit how much cyber skill transfers, even when the underlying task is one its own rules allow.

The worry is dual-use. A model that helps defenders find software flaws can also help attackers exploit them; one that assists with biological research could help someone develop a harmful pathogen. A distilled copy that learned the capability without the refusals is, from OpenAI’s point of view, a safety problem first and a business problem second.

The Pioneer Building in San Francisco, which housed OpenAI’s offices in 2019
Photo: the Pioneer Building in San Francisco, which housed OpenAI’s offices in 2019. HaeB, CC BY-SA 4.0, via Wikimedia Commons.

Three layers of defense#

The defenses work at two levels. At the request level, OpenAI trains models to recognise extraction attempts and refuse them, runs classifiers over inputs, outputs and tool-call arguments to catch leaks — including reasoning smuggled out through a tool’s arguments rather than the final answer — and applies API controls that limit how protected reasoning can be reused.

The most consequential API change is structural. In GPT-6.2 Astra and GPT-6.2 Sol, a model’s protected reasoning is tied to the conversation history that produced it: new messages can build on earlier reasoning, but if an earlier message is rewritten, the API drops the reasoning that followed it. Older model families with weaker safeguards are also prevented from reading GPT-6.2’s protected reasoning. The change closes a pathway described in a September disclosure, in which attackers moved encrypted reasoning between conversations and got the model to reveal it — including, OpenAI says, a pathway that could have exposed another user’s encrypted reasoning.

At the campaign level, OpenAI looks for patterns across requests and accounts that reveal coordinated extraction, enforces against the networks it finds, and shares indicators of compromise with other model providers. It also applies a “consistent-standards” policy across the partners that deploy its models, so the same safeguard floor holds wherever the models run.

Naming names — and asking Congress for help#

The post’s sharpest turn is the attribution. OpenAI says that in February it briefed Congress on activity indicating DeepSeek was attempting to distill U.S. frontier models, and that in September it disrupted a campaign to extract its models’ protected reasoning, attributing a core part of that activity to individuals associated with Moonshot AI, the developer of Kimi. It says it plans to share more soon about recent campaigns it has disrupted.

Weekly enforcement against these campaigns, the company says, has ranged from thousands to tens of thousands of accounts, “with the rate increasing significantly this past summer.” The post also turns into a policy document: OpenAI says it supports the Chip Security Act, the AI OVERWATCH Act, and the MATCH Act, arguing that stronger export controls limit adversaries’ access to training hardware while distillation defenses make it harder to acquire the same capabilities from models already trained.

What to watch#

Four things will decide whether this amounts to more than a strong press release. First, independent testing of the 0.2% figure — no outside lab has yet reproduced OpenAI’s red-team results, and the company has not named the six attacker models. Second, whether the defenses hold up as attackers adapt; OpenAI’s own framing is a cat-and-mouse game, and its adaptive evaluations are designed to find the next weakness. Third, whether DeepSeek or Moonshot answer the attributions publicly — Moonshot, for its part, stripped cyber material from Kimi K3.1 Pro’s training data before publishing its weights in October, according to one account. Fourth, whether Congress acts on the three bills OpenAI backs.

One tension runs underneath all of it. The same wall that keeps a model’s reasoning away from attackers also keeps it away from the developers and users who act on its answers — and it protects OpenAI’s commercial lead at the same time as it protects its safeguards. Whether the industry treats that as a safety standard or a moat will shape how far other labs follow.

Sources: OpenAI, “Combatting adversarial distillation”; AI Weekly, “OpenAI says GPT-6.2 cuts reasoning-extraction attacks to 0.2%”; AIVIO, “OpenAI Cuts Reasoning-Extraction Attacks to 0.2%”