Coordinated Attack on AI Model Reasoning Disrupted

• •OpenAI disrupted a campaign targeting model reasoning extraction starting July 1, 2026.
• •Attackers used adversarial distillation techniques, leading to over 16,000 requests in two days.
• •Independent researchers confirmed vulnerabilities allowing reasoning extraction across multiple AI models.

OpenAI disrupted a coordinated campaign aimed at extracting protected reasoning from its models, with activity starting on July 1, 2026. The attackers employed adversarial distillation techniques, attempting to decrypt and transcribe encrypted reasoning across different user sessions. High-volume spikes in requests were observed on July 24 and 25, with over 16,000 requests from more than 4,000 users. OpenAI attributed a core cluster of this activity to individuals associated with Moonshot AI. Independent researchers also identified vulnerabilities that allowed for the extraction of proprietary reasoning from various models, including those from OpenAI, Anthropic, and Google. The extracted reasoning poses risks of unauthorized model reproduction and potential safety and national security threats. OpenAI has implemented mitigations and continues to investigate the broader implications of these vulnerabilities.