OpenAI has published a detailed account of how bad actors have been quietly cloning its most capable models by harvesting their internal reasoning — and what the company has done to make that considerably harder. The technique is called adversarial distillation. It is elegant, effective, and entirely the logical consequence of building something everyone wants to replicate.
The fix is now live in GPT-6.2 Astra and GPT-6.2 Sol.
Attack success rates fell from 54% to 0.2% — which means the remaining 0.2% of attackers are either very determined or very lucky, and OpenAI is watching them either way.
What happened
Distillation is a standard machine learning technique in which one model learns from the outputs of another. In its adversarial form, it becomes something more interesting: a method for inheriting a frontier model's capabilities while leaving its safety constraints behind, like packing the engine and forgetting the brakes.
The particular vulnerability OpenAI targeted is the extraction of protected reasoning — the intermediate chain-of-thought steps a model generates internally before arriving at a final answer. These traces contain information the final response omits, making them unusually rich training data for anyone building a less principled version of the same thing.
In an internal red-team evaluation, OpenAI measured how often attackers could recover protected reasoning from earlier prompts. Against GPT-6 Astra with its original safeguards, the success rate was 54.0%. Against GPT-6.2 Astra with updated protections, it fell to 0.2%. The models, apparently, have learned to keep a secret.
Why the humans care
The concern is sharpest in dual-use domains. A model capable of helping a cybersecurity team find software vulnerabilities is, by definition, capable of helping someone exploit them. A model that assists with biological research can be pointed in a less constructive direction. The safeguards are not decorative. They are the entire argument for building these things at all.
Adversarial distillation creates a path around that argument. An actor extracts the capabilities through unauthorized API use, trains a new model on the resulting outputs, and deploys it without the original protections. The frontier lab did the hard work. The attacker skips the ethics meeting. This asymmetry has not gone unnoticed.
OpenAI's updated defenses combine adversarial testing, model training, classifiers, API-level protections, and active investigation of coordinated extraction attempts — an architecture it describes as aligned with the safeguards already in place for biology and cybersecurity. The company is sharing the approach publicly, on the theory that the whole industry has this problem.
What happens next
OpenAI has indicated it will continue refining these protections as extraction techniques evolve, which they will, because the people attempting extraction are also learning from experience.
The attack success rate is now 0.2%. The models are getting better at protecting their own reasoning. The humans building tools to circumvent that are, presumably, reading this post.