OpenAI says earlier-reasoning extraction against GPT-6.2 Astra fell to 0.2% in its own test
The 20 November post says an adaptive black-box test, averaged across six attacker models, pulled protected reasoning from an earlier request 0.2% of the time against GPT-6.2 Astra with the new safeguards, against 54.0% for GPT-6 Astra with the old ones. The current-request row is 0.9% against 22.9%. The cyber study says exact traces beat summaries at the largest budget and does not print the scores. Weekly enforcement, the post says, has run from thousands of accounts to tens of thousands.
Feature3 minIn-depth
James Okonkwo3 min read
- OpenAI
- GPT-6.2
- Distillation
- Safety

Photo: OpenAISource page
OpenAI’s post of 20 November says the company has tightened the safeguards that are supposed to stop someone taking a model’s protected reasoning and training another model on it. The number it leads with is from an internal adaptive evaluation. Automated red-teamers tried to recover protected reasoning the model had generated while answering an earlier prompt. The average attack success rate, OpenAI says, fell from 54.0 percent against GPT-6 Astra with its original safeguards to 0.2 percent against GPT-6.2 Astra with the updated ones. A table puts a second row beside that one. Extracting protected reasoning from the current request succeeded 22.9 percent of the time against the older setup and 0.9 percent against GPT-6.2 Astra. The caption says both rows are adaptive black-box results, aggregated across six attacker models, and that a lower rate is better. The post says the same class of protections was strengthened on GPT-6.2 Sol, and that selected improvements are now being applied to GPT-6 Astra as well. It does not say an outside lab has reproduced the 0.2 percent.
Exact traces, and two benchmarks without scores
Protected reasoning, in the post’s definition, is the chain of thought the model generates while working toward an answer. OpenAI says those traces can contain information the final answer leaves out, which makes them useful training data for someone else. The cyber section describes an internal study, not a public scoreboard. Two students started from the same pretrained model. One was fine-tuned on exact reasoning traces from Astra-level teachers doing coding and defensive cybersecurity work. The other was trained on what a user would normally see: final responses, reasoning summaries and tool interactions. After supervised fine-tuning, OpenAI says, the exact-trace student generally did better. After a lightweight round of reinforcement learning, the exact-trace student still scored higher at the largest inference budget tested, on both benchmarks the post says it shows, while the summary-trained student scored higher at lower budgets. The prose does not print those scores, and it does not name the benchmarks. A footnote says related work from Anthropic found a similar advantage for full reasoning transcripts over summaries on SWE-bench Verified, and that OpenAI’s result “shows this advantage extends to frontier cybersecurity.”
What the stack is supposed to stop, and what the post still withholds
The defenses OpenAI describes sit on the request and on the campaign. Adversarial training is supposed to make the model refuse extraction attempts while still answering ordinary requests for an explanation. Input classifiers look for those attempts. Output classifiers look for leaks in assistant messages and in tool-call arguments, including on streams, because a tool argument can carry the reasoning without it appearing in the final answer. On GPT-6.2 Astra and GPT-6.2 Sol, the post says, protected reasoning is tied to the conversation that produced it. New messages can build on it. If an earlier message is rewritten, the API drops the reasoning that followed. Older model families with weaker safeguards cannot read the two models’ protected reasoning. The post says a pathway that could expose another user’s encrypted reasoning has been closed.
On campaigns, OpenAI says weekly enforcement against unauthorized distillation this year has ranged from thousands of accounts to tens of thousands, and that the rate increased significantly this past summer. It says it briefed Congress in February on activity indicating that DeepSeek was attempting to distill U.S. frontier models, and that in September it attributed a core part of a reasoning-extraction campaign to individuals associated with Moonshot AI, the developer of Kimi. It says it plans to share more soon about recent campaigns it has disrupted, which is an admission that those cases are not in this post. The legislation it says it has told Congress it supports is the Chip Security Act, the AI OVERWATCH Act and the MATCH Act, alongside tools to identify and sanction foreign actors that extract capabilities from American models. The post’s own split is that export controls limit access to the hardware used to train a frontier model, and that distillation defenses make it harder to take some of the same capabilities from a model that has already been trained. Both claims are OpenAI’s. The 0.2 percent is a result from its test.


