AI · Nov 9, 2026
OpenAI’s Control Room puts a fail-closed kill switch inside the inference engineOpenAI treats chain-of-thought as the crown jewels, and asks Congress to help guard them
· Updated · Signal Desk newsroom
Analysis of reporting by OpenAI. How we work

OpenAI’s new anti-distillation playbook says leaked reasoning traces teach copycat models real cyber skills. The defenses look serious; the policy pitch blends safety with competitive self-interest.
Key takeaways
- OpenAI says attack success at extracting protected reasoning fell from 54.0% on GPT-6 Astra’s original safeguards to 0.2% on GPT-6.2 Astra in an internal adaptive test.
- An internal study found students trained on exact reasoning traces generally beat students trained on summaries at cyber and coding tasks.
- Weekly enforcement against distillation campaigns has hit thousands to tens of thousands of accounts, rising sharply this summer.
- GPT-6.2 Astra and Sol now tie reasoning to the conversation that produced it and block older model families from reading it.
- OpenAI backs the Chip Security Act, the AI OVERWATCH Act and the MATCH Act, and points to DeepSeek and Moonshot AI activity.
From IP complaint to safety argument
Labs have grumbled about distillation for years, usually in the language of stolen work. OpenAI’s new post reframes it as a safety problem. The argument is straightforward: a rival that trains on a frontier model’s outputs can inherit its capabilities without inheriting the refusals and classifiers wrapped around them. A model that will not help with an attack can teach a model that will.
That framing is not just rhetoric. OpenAI describes an internal experiment in which a student model learned from Astra-level teachers working on coding and defensive security tasks. Students given the exact reasoning traces generally outperformed students given only the summaries users normally see, at least after fine-tuning; after a round of reinforcement learning the trace-trained student kept its lead only at the largest inference budget. Even benign, allowed work, it turns out, can transfer offensive skill when the full chain of thought leaks.
What the defenses actually do
The engineering is layered. Models are adversarially trained to refuse extraction attempts while still explaining themselves to ordinary users. Input classifiers look for extraction attempts and output classifiers scan both messages and tool-call arguments, which matters because a model can leak reasoning into a tool call without ever showing it in the answer. At the API level, GPT-6.2 Astra and GPT-6.2 Sol bind protected reasoning to the conversation that produced it: rewrite an earlier message and the reasoning after it is dropped. Older model families with weaker safeguards cannot read the new models’ reasoning at all.
The headline number is striking. In an adaptive red-team evaluation, average attack success fell from 54.0% against GPT-6 Astra with its original protections to 0.2% against GPT-6.2 Astra. Those are OpenAI’s own tests with OpenAI’s own attackers, so they deserve the usual caution, but the company at least describes static replays, adaptive attackers and white-box variants rather than a single flattering benchmark. Some fixes are being backported to GPT-6 Astra.
The campaigns are bigger than the last disclosure suggested
In September OpenAI described a coordinated extraction campaign it attributed in part to people associated with Moonshot AI. This update widens the lens: multiple campaigns this year, with weekly enforcement actions ranging from thousands to tens of thousands of accounts and a sharp rise over the summer. OpenAI also says it briefed Congress in February on activity indicating DeepSeek was trying to distill U.S. frontier models, and promises more disclosures soon.
One detail deserves more attention than it will get. OpenAI says it closed a pathway that could have exposed another user’s encrypted reasoning. That is a privacy fix, not only an anti-distillation one, and it shows how a defense aimed at foreign copycats also hardens the product for ordinary customers.
Where safety and moat overlap
Then comes the policy ask. OpenAI supports legislation to identify and sanction foreign actors that extract capabilities from American models, plus the Chip Security Act, the AI OVERWATCH Act and the MATCH Act. It argues that export controls and distillation defenses together preserve a U.S. lead and give developers room to slow releases for safety testing.
There is a real case here. Cheap, unguarded copies of frontier cyber capability are a genuine risk. But hidden reasoning is also the clearest competitive moat a lab has, and every argument for guarding it doubles as an argument against transparency. Researchers and customers increasingly want to inspect how models reason. The more chain-of-thought becomes a protected asset, the harder that gets. OpenAI’s recommendations to other labs, such as keeping attack logs as regression tests and coordinating enforcement with deployment partners, are useful. The industry should adopt the defenses while being wary of letting safety language decide who gets to see inside the box.
Who feels it
- Developers on OpenAI’s API
- Workflows that edit earlier messages or move reasoning across models will lose access to prior reasoning by design.
- Rival labs
- Training on OpenAI outputs gets costlier and riskier, especially for teams named in government briefings.
- Security researchers
- Stronger controls on reasoning traces may limit outside study of model behavior even as they reduce leakage.
- Policymakers
- OpenAI is offering a ready-made bill list that pairs chip controls with sanctions for model extraction.
What to watch
- The additional campaign disclosures OpenAI says are coming soon, and whether they name new actors.
- Movement on the Chip Security Act, AI OVERWATCH Act and MATCH Act in Congress.
- Whether Anthropic, Google and others publish comparable extraction-resistance numbers.
- Independent red-team results that test OpenAI’s 0.2% attack-success figure.
FAQ
What is adversarial distillation?
It is using a model’s outputs without authorization to train another model with similar capabilities, which can then be deployed without the original safeguards.
Why does OpenAI focus on reasoning traces?
Its internal study found students trained on exact chains of thought generally learned more cyber capability than those trained on summaries, because the traces contain information final answers leave out.
How effective are the new safeguards?
In OpenAI’s adaptive red-team test, average attack success at extracting protected reasoning fell from 54.0% on GPT-6 Astra to 0.2% on GPT-6.2 Astra. These are internal results that have not been independently replicated.
Companies: OpenAI
More on OpenAI
AI · Nov 11, 2026
OpenAI’s $2 million for children’s AI safety research matters for one promise: publishing bad newsAI · Nov 17, 2026
OpenAI says its models now shave area off production chips, including AMD’s next flagshipsAI · Nov 3, 2026
OpenAI’s GPT-5.6 Cyber reaches AWS GovCloud through Daybreak Red, with two approval gatesAI · Nov 5, 2026
Perplexity Teammates bets that always-on work agents are worth $200 a month