🛡️OpenAI Cuts Distillation Attack Success From 54% to 0.2% in New Safeguards
Copying frontier models just got a lot harder
TL;DR
OpenAI reports its updated safeguards cut the average attack success rate against adversarial distillation from 54.0% on GPT-6 Astra to 0.2% on GPT-6.2 Astra. Unauthorized distillation can copy frontier capabilities, including risky cyber skills, so if you ship on top of these APIs, expect tighter reasoning protections.
OpenAI says it has dropped the success rate of adversarial distillation attacks from 54.0% to 0.2% in a single model generation. That's the jump from GPT-6 Astra with its original safeguards to GPT-6.2 Astra with updated ones. Distillation itself is a normal technique where one model learns from another's outputs. The problem is when someone does it without authorization to clone a frontier model's abilities without the safety layers attached. Why should you care? If you build security tooling, the same model that finds vulnerabilities for defenders can help attackers exploit them. A student model trained on a teacher's benign outputs can pick up offensive cyber skills it was never meant to have. OpenAI's defense stacks adversarial training, input and output classifiers, and API-level binding of reasoning traces to conversation history, plus campaign-level enforcement. The numbers matter here. In an internal study, students trained on exact reasoning traces generally beat students trained on summaries, which is why protecting those traces is the lever. Expect this to shape how the API treats chain-of-thought-style output going forward. What to watch: whether these protections hold up against determined extraction campaigns, and how much they constrain legitimate users who depend on detailed reasoning output.

Key Points
OpenAI reports average attack success fell from 54.0% against GPT-6 Astra with original safeguards to just 0.2% against GPT-6.2 Astra with updated safeguards, a dramatic single-generation improvement.
Keep reading — it's free
Enter your email to keep reading — plus the best of AI & tech, daily. Free, forever.
Already a member? Sign in
Comments
Be the first to comment