2026-11-20 · ← News
OpenAI cuts reasoning extraction success from 54% to 0.2%
OpenAI says GPT-6.2 Astra reduced the success rate of an adaptive attack on earlier protected reasoning from 54% to 0.2%. The result comes from an internal test and protects a valuable model trace, not every route to copying its capabilities.
The image could not be loaded.
OpenAI has described its defence against adversarial distillation, where an attacker collects another model's outputs and trains a new system on them. Its headline result concerns protected reasoning: in an internal adaptive black-box test, extraction success for a trace from an earlier request fell from 54% on GPT-6 Astra to 0.2% on GPT-6.2 Astra.
GPT-6.2 Astra closes the most direct route to internal traces
OpenAI used six attacker models that adjusted their approach after seeing the system's responses. For extracting reasoning from the current request, average success fell from 22.9% to 0.9%. The company says the protections also cover GPT-6.2 Sol.
The defence combines model training, input and output classifiers, and API rules. It also inspects tool-call arguments because a model can leak a sensitive trace to a tool without writing it to the user. Protected reasoning is bound to conversation history, while older model families are blocked from reading traces created by the new models.
The distillation fight now spans campaigns rather than single accounts
OpenAI says weekly enforcement against unauthorised distillation has involved thousands to tens of thousands of accounts. A single-request filter is therefore insufficient. The company looks for patterns across accounts, feeds investigations back into training and shares indicators with other providers.
For API operators, the issue extends beyond intellectual property. Data that teaches a smaller model to code or solve a biology task may transfer a capability whose use the original system restricted. Without campaign detection, even a strong filter is merely a turnstile in front of a crowd of disposable accounts.
An internal eval measures trace leakage, not the end of capability copying
OpenAI designed and reported the evaluation itself. It compares two model generations with different safeguards, so it does not isolate model changes from the rest of the protection stack. Nor does it show how much capability can be learned from final answers, reasoning summaries and tool calls without recovering an exact chain of thought.
Independent research also suggests that approximate traces can regain value after further reinforcement learning. That supports layered defence, but it makes 0.2% a narrow security result rather than a declaration of victory.
Robustness across new attacks and production scale will settle the claim
The next useful evidence is testing with different student models, training recipes and query budgets. It will also matter whether the safeguards degrade legitimate explanations or long-running agent workflows that rely on reasoning summaries and tool calls.
OpenAI says it will disclose more about disrupted campaigns. Detection time, extracted data volume and results outside its own evaluation would reveal whether the system raises the real cost of distillation or mainly locks one specific opening.
Lilith's verdict
OpenAI showed a lock that resisted its own picks in 99.8% of attempts. A capability thief may still carry the house away brick by brick through answers, summaries and tool calls.