A model’s value is not just in its answers but in the steps it takes to get there. OpenAI has been focused on stopping attackers from extracting those steps—its protected reasoning or chain-of-thought traces—by distilling outputs from advanced models like GPT-6.2 Astra and Sol. The risk: if bad actors can collect and use those traces, they can train comparable models without the original safeguards, potentially unleashing offensive cyber capabilities the original model would refuse to provide.
OpenAI now reports a sharp drop in attackers’ ability to extract these traces. In internal adaptive evaluations using strong automated red-teaming, the attack success rate against GPT-6 Astra fell from 54.0% with its original safeguards to just 0.2% in GPT-6.2 Astra after the new defences were applied.
Why distillation matters for safety
Distillation lets one model (the student) learn from another (the teacher), often by training on the teacher’s outputs. When applied adversarially, this can transfer advanced capabilities—especially in dual-use domains like cybersecurity—without transferring the original model’s safeguards. For example, a model that helps defenders find vulnerabilities could be distilled into one that helps attackers exploit them, if its reasoning traces are exposed.
How reasoning traces drive capability transfer
OpenAI’s internal study compared training student models on either final answers, reasoning summaries, or full reasoning traces from Astra-level teachers. The student trained on exact reasoning traces outperformed the one trained on summaries, especially after reinforcement learning, suggesting that these traces are a crucial vector for capability transfer. Protecting them is now central to OpenAI’s approach to cyber safety.
Layered defences: request-level and campaign-level
The new safeguards work at multiple levels:
- Request-level: Model behaviour checks, classifiers, and API controls now limit extraction of reasoning traces.




