OpenAI publishes adversarial distillation defense and evaluation results
Understand this story
OpenAI has published a set of methods to detect and block adversarial distillation, saying it combines adversarial testing, model training, classifiers, and API-layer protections, and strengthens protection of models' protected reasoning (chain-of-thought). OpenAI's internal evaluation says the success rate of reasoning-trace extraction attacks against GPT-6.2 Astra dropped to 0.2%. OpenAI said some of the protection improvements have also been applied to GPT-6 Astra, and that it blocks older, less-protected model families from reading the protected reasoning of GPT-6.2 Astra and Sol. OpenAI said it will later disclose related actions it has recently blocked.
AI overview based on source reports · Updated 32 min ago
Coverage timeline
Follow the coverage to explore different perspectives.
- OpenAI NewsFeaturedCombatting 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.
Activity for this story
Not enough continuous observations to show a trend yet.