OpenAI says GPT-6.2 Astra and Sol are substantially harder to attack for protected internal reasoning. In adaptive black-box tests, extraction from prior requests succeeded 0.2% of the time against GPT-6.2 Astra, down from 54% for GPT-6 Astra. Attacks targeting an active request fell from 22.9% to 0.9%.

The company combines targeted training with filters on messages, outputs and tool calls. API restrictions also prevent older models from reading protected reasoning. OpenAI says it is transferring selected improvements back to GPT-6 Astra and bans thousands to tens of thousands of accounts each week when it identifies coordinated model-distillation campaigns.

These are OpenAI’s internal red-team measurements, not a guarantee that every extraction technique has been eliminated. Attack success also depends on the adversary models and test setup used. Anthropic has described similar pressure and large-scale account bans, suggesting that reasoning theft is becoming an operational security problem across frontier labs. The sharp measured reduction raises the cost of known attacks, but continued adaptive testing will be needed as attackers change tactics.