OpenAI says it identified and disrupted a coordinated campaign designed to extract protected model reasoning at scale. The company attributed a core cluster of the activity to people associated with Moonshot AI, while noting that the activity involved model interactions rather than a breach of encryption, databases or stored user conversations.
OpenAI described the technique as adversarial distillation: systematically eliciting outputs to help train, reproduce or improve another model. The company said the campaign spiked to 16,000 attempted requests across more than 4,000 users on two days in July and that it later found related prompt-pattern activity across more than 15,000 users.
Security implications
The case frames model extraction as an AI-security concern that extends beyond traditional account compromise. Providers need abuse detection, rate and pattern controls, and mechanisms to contain outputs that could expose protected reasoning. Organizations consuming AI services should also distinguish provider claims and attribution from independently verified technical evidence.
Source: The Hacker News report; OpenAI announcement.