AI NewsModels & agentsAnnouncement
OpenAI says model-copying attack rate fell from 54 percent to 0.2 percent
OpenAI says attacks that copy its model capabilities fell from 54 percent to 0.2 percent on GPT-6.2 Astra.

Why it mattersA team building on a vendor model can expect fewer copycat forks of that model to appear in cheap wrappers.
A team building a product on a vendor model worries that a competitor will copy that model for pennies. OpenAI said on Thursday that it has cut that risk on its newest models. In a post on its blog, the company said the attack success rate fell from 54 percent to 0.2 percent. The old rate is on GPT-6 Astra. The new rate is on GPT-6.2 Astra. The figure is OpenAI's own.
OpenAI defines adversarial distillation as a copycat training a cheaper model on the outputs of a frontier one. The attacker skips the safety work. The copy then goes to users without the same guardrails. OpenAI says the risk is biggest for cybersecurity and biology, where a helper for defenders can also help attackers.
What OpenAI changed
The post names three defences, used together. The company trained the model to resist the copying. It added new classifiers around the API. And it tightened how the model shares the reasoning steps it writes to itself. OpenAI calls the last part "protected reasoning". The test that fell from 54 percent to 0.2 percent is OpenAI's own. The post leaves out two details a reader would want: the attack technique used, and how many prompts the rate was measured over.
OpenAI also says exact copies of reasoning traces transfer more capability than summaries do. It says this was an internal study. No outside paper is cited. The company adds that it has banned accounts used for these attacks. It says the counts range from "thousands to tens of thousands" per action.
How to read the number
The headline figure is a 270-fold drop. That is a company's statement about its own model on a test it designed and ran. It is not a statement any independent lab has verified. Vendor safety numbers often fall sharply the first time the vendor measures them. The numbers rise again when a third party retests on a wider set of attacks. On its own, 0.2 percent is a starting figure, and the real one may be higher.
A team building on GPT-6.2 Astra does not pick a model on this number alone. Pick it on three. Does the model do your work as well as the model you use now, measured by your own eval. Does the token bill come out the same or lower on real traffic. And does the safety team at your vendor publish a figure you can check with your own red team. OpenAI's post says what the company did and what it measured. It stops there.
Source
This item was written by an AI system from the linked source. Reveneau is responsible for what it publishes.
Get AI News in your inbox
New developer tools, model and agent releases, and how teams are actually using them to release software. Short, and only when there is something worth reading.


