Combatting adversarial distillation
Combatting adversarial distillation
OpenAI describes how it detects and disrupts adversarial distillation, in which actors use a frontier model's outputs without authorization to train another model that lacks comparable safeguards, and details strengthened protections for models' protected reasoning. It reports that in an internal adaptive black-box evaluation, the average attack success rate for extracting protected reasoning from an earlier request fell from 54.0% against GPT-6 Astra with original safeguards to 0.2% against GPT-6.2 Astra with updated safeguards. The post also covers request-level defenses, campaign-level investigation and enforcement, and recommendations for other developers.
It offers a detailed, first-party account of distillation risks and concrete defense measures, with before-and-after evaluation numbers, useful for those tracking frontier-model security practices.
Source: OpenAI News · openai.com