OpenAI says it shut down a coordinated effort to copy how its models think.
The company disclosed that it disrupted a campaign aimed at extracting the internal reasoning of its models, a technique known as distillation, where an attacker feeds a target model prompts and uses the outputs to train a cheaper copycat. OpenAI did not name who ran the campaign or how many models or accounts were involved. It said the operation was coordinated, implying multiple accounts or automated scripts working together rather than one person probing the API. In response, OpenAI says it is strengthening its defenses against this kind of adversarial distillation going forward.
Reasoning traces, the step-by-step thinking a model produces before an answer, are increasingly the expensive part to build and the easiest part to steal if a rival just scrapes enough of them. For a company that charges a premium for reasoning models, blocking that leakage is as much a business decision as a security one.
OpenAI's post is light on specifics, naming no attacker and no numbers, so take the coordinated-campaign framing as the company's side of the story until someone corroborates it.