OpenAI says the success rate of adversarial distillation attacks designed to extract protected reasoning from GPT-6.2 Astra has fallen to 0.2%, down from 54.0% with GPT-6 Astra. For attempts targeting reasoning linked to the current request, the rate dropped from 22.9% to 0.9%. The company combines adversarial training with API-level measures that tie reasoning to specific conversation histories. Throughout 2026, OpenAI says it has taken weekly enforcement actions against thousands of accounts and shared indicators of compromise with other model providers. Published following the release of GPT-6.2 Astra and GPT-6.2 Sol, these protections aim to reduce risks in areas such as cybersecurity and biological research.
Source: Read the original article


