OpenAI says it has strengthened protections against attempts to extract private reasoning from GPT-6.2 Astra and GPT-6.2 Sol. It also outlined how it detects coordinated efforts to use its models’ outputs to train other models without authorization.
OpenAI calls this activity adversarial distillation. A model trained on another model’s answers or reasoning could gain useful capabilities without inheriting its safeguards, the company says. OpenAI is particularly concerned about cybersecurity and biology, where capabilities can have both beneficial and harmful uses.
Extraction attack success dropped below one percent
In an internal test using automated attackers, the average success rate for recovering reasoning generated during an earlier prompt fell from 54.0% against GPT-6 Astra with its original safeguards to 0.2% against GPT-6.2 Astra with updated safeguards. OpenAI says separate adaptive tests found success rates below 1% for two types of reasoning extraction attack against GPT-6.2 Astra.
The protections include model training, checks on incoming requests and outgoing responses, and API controls. OpenAI says its output checks cover tool-call arguments as well as assistant messages, including streaming responses. In GPT-6.2 Astra and Sol, protected reasoning is tied to the conversation history that produced it. If an earlier message is rewritten, the API drops the reasoning that followed it.

OpenAI takes weekly action against account networks
OpenAI says it also investigates coordinated extraction attempts and takes action against networks of accounts when it identifies violations. Its weekly enforcement actions against such campaigns have ranged from thousands to tens of thousands of accounts this year. The company says it shares indicators of compromise with other model providers.
OpenAI is now applying selected safeguard improvements to GPT-6 Astra. It says it plans to share more about recent campaigns it has disrupted.



