Combatting adversarial distillation
- Published
- Record updated
Summary
OpenAI describes a layered defense against adversarial distillation, in which an actor uses a frontier model's outputs without authorization to train a second model and deploy it without comparable safeguards. The company says it strengthened protections for protected reasoning, the intermediate chain-of-thought traces, in GPT-6.2 Astra and GPT-6.2 Sol. In an internal adaptive evaluation, the average attack success rate against recovering protected reasoning fell from 54.0% against GPT-6 Astra to 0.2% against GPT-6.2 Astra.
Mitigation
Strengthened defenses against protected reasoning extraction in GPT-6.2 Astra and GPT-6.2 Sol, with updated safeguards that align with those used for biology and cybersecurity. The approach combines adversarial testing, model training, classifiers, API-level protections, and investigations into coordinated extraction attempts.
Related items
- InfoIn a year of AI solving maths problems, Anthropic may have just solved the biggestSame vendor · BBC Technology
- InfoAI R&D-relevant model capabilities are improving rapidly and are getting harder to measureSame vendor · METR
- InfoSoftBank plunges over 16% as AI sell-off hits Asian tech stocks following Anthropic debutSame vendor · CNBC Technology
- InfoGPT-6.2 Astra vs. Sol: Features, Pricing & Security | StraikerSame vendor · Straiker Blog
- InfoCan AI help scientists identify promising drug candidates?Same vendor · OpenAI Blog