OpenAI says it strengthened reasoning safeguards in Astra and Sol

The company says internal tests found extraction success fell from 54.0% to 0.2% for earlier requests and from 22.9% to 0.9% for current requests, comparing different Astra generations and safeguards.
In this article
OpenAI says it has strengthened protections against the extraction of private reasoning from its GPT-6.2 Astra and GPT-6.2 Sol models. In a November 20 disclosure, the company, the only source for this account, said the measures are intended to stop others from using that reasoning to reproduce model capabilities without comparable safeguards. OpenAI’s disclosure describes internal tests of the Astra protections.
The practice OpenAI calls adversarial distillation uses one model’s outputs without authorization to train another. OpenAI says the resulting model can acquire capabilities without inheriting the original model’s safeguards, potentially allowing uses the original would refuse.
What the internal tests showed
Protected reasoning means the intermediate steps a model generates while working toward an answer. OpenAI says those private traces can contain information missing from the final answer; users normally see responses, reasoning summaries and interactions with tools instead.
OpenAI says its automated attackers adjusted their methods in response to the system and had the same view as an external user. The company aggregated the results across six attacker models in a test setup it says closely mirrors production.
| Reasoning targeted | GPT-6 Astra, original safeguards | GPT-6.2 Astra, updated safeguards |
|---|---|---|
| Earlier request | 54.0% | 0.2% |
| Current request | 22.9% | 0.9% |
The comparison changes both the model generation and the safeguards, so it does not isolate the effect of the safeguards alone. OpenAI did not give a comparable result for Sol or independent or live-production validation of the rates.
In a separate internal cybersecurity study, OpenAI says a model trained on exact reasoning traces generally performed better than one trained on summaries after supervised training. After further training, the exact-trace model scored higher on both benchmarks shown at the largest inference computing budget tested; the summary-trained model scored higher at lower computing budgets.
How the protections work
OpenAI says it trains models to refuse attempts to obtain protected reasoning while still providing legitimate explanations. It also uses classifiers—systems that identify specified kinds of content—to detect extraction attempts in inputs and block suspected leaks in responses or instructions sent to tools.
The company ties the newer models’ protected reasoning to the conversation that produced it. Developers using its application programming interface, or API, can add messages and let a model build on earlier reasoning; if they rewrite an earlier message, the API drops the reasoning that followed it. Older model families cannot read the newer models’ protected reasoning, OpenAI says.
Campaigns and next steps
OpenAI says it briefed Congress in February on activity indicating that DeepSeek was attempting to distill US models. In September, it said it had disrupted a reasoning-extraction campaign and attributed a core part of the activity to people associated with Moonshot AI, the developer of Kimi.
The company says it investigates coordinated campaigns and takes action against account networks when it identifies violations. Weekly enforcement has ranged from thousands to tens of thousands of accounts, with the rate increasing significantly over the summer. OpenAI says it coordinates standards and enforcement with deployment partners and shares indicators of compromise with other model providers.
OpenAI says it is applying selected improvements to the older GPT-6 Astra and plans further study of capability transfer across student models and training methods. It also plans to share more about recently disrupted campaigns, without specifying a date.
Follow AveNews
Earlier in Technology
SoftwareBrickLink targets November 24 return after marketplace outage
The LEGO resale site went offline on 19 November while its team investigated unusual activity. Its reopening target depends on the investigation and work to restore service safely.
2 min
RoboticsWaymo issues two robotaxi recalls over freeway incidents
Alphabet’s self-driving unit cites incidents involving closed lanes and manual traffic control. The recalls follow California’s restriction on driverless freeway operations and the company’s pause in its other US markets.
2 min
RightsYouTube reaches tentative $3 billion US schools deal, sources say
The proposed settlement includes safeguards for teens, but a majority of districts must sign on and the judge overseeing the litigation must approve it.
2 min
