OpenAI Adversarial Distillation: GPT-6.2 Safeguards (Nov 20)
Reported by explainx.ai, The Decoder and Unite.AI3 sources ↓
OpenAI says adaptive attacks extracting protected reasoning from earlier requests succeeded 0.2% of the time on GPT-6.2 Astra, down from 54% on GPT-6 Astra. For active requests, the rate fell from 22.9% to 0.9%.
Key points
- Selected safeguard improvements are also being applied to GPT-6 Astra
- The evaluation combined adaptive black-box attacks from six attacker models
- OpenAI says it bans thousands to tens of thousands of accounts weekly over unauthorized distillation campaigns
Sources
Details
The tests measured attacks that adapt to model responses while seeing only what an external user would see. The figures are averages across six attacker models, according to OpenAI; the company describes the evaluation as internal automated red-teaming.
OpenAI says request-level protections include training models to recognize extraction attempts, filters for incoming and outgoing messages, and checks of tool-call arguments. The updated API also ties protected reasoning to the conversation history that produced it, and prevents older model families from reading GPT-6.2 Astra and Sol’s protected reasoning.
For coordinated campaigns, OpenAI says it uses request and account signals to investigate suspected extraction and uses discovered attacks to train and test safeguards. The company reports weekly enforcement against thousands to tens of thousands of accounts in 2026, with enforcement increasing significantly during the summer.
Written from coverage as of
Images
- The Decoder
- Unite.AI