OpenAI unveiled multi-layered defenses against adversarial distillation, cutting protected chain-of-thought extraction on GPT-6.2 Astra from 54.0% to 0.2% for prior requests and from 22.9% to 0.9% for current ones. The measures combine adversarial training, input/output classifiers and API-level reasoning isolation, backed by weekly enforcement against thousands to tens of thousands of violating accounts.
Adversarial Distillation
对抗性蒸馏
Unauthorized training on frontier model outputs to transfer capabilities while bypassing original safeguards.
Curated Intelligence
OpenAI has sharply hardened reasoning-extraction defenses for GPT-6.2 Astra and Sol, cutting attack success rates on prior requests from 54.0% to 0.2% and on active requests from 22.9% to 0.9% in internal automated tests. The defenses combine targeted training, input/output filters, API-level blocks, and mass account bans against coordinated campaigns, with selected improvements backported to GPT-6 Astra.
OpenAI outlined how it detects and disrupts adversarial distillation by combining adversarial testing, model training, classifiers and API-level protections, with a focus on hardening protected reasoning. The company warns that actors can use a frontier model's outputs without authorization to train a similarly capable model and deploy it without comparable safeguards, for example distilling cyber capabilities for attacks the original model would refuse. OpenAI also shared lessons intended to strengthen defenses industry-wide.
OpenAI outlines a defense framework against adversarial distillation that combines adversarial testing, model training, classifiers, and API-level protections to stop attackers from training unsafeguarded clones on frontier model outputs. It also details hardened protections for protected reasoning and notes that full reasoning transcripts outperform summaries for distillation on SWE-bench Verified and frontier cybersecurity tasks, offering lessons for industry-wide safeguards.