OpenAI Blocks 15,000 Requests Trying to Extract Protected Model Reasoning

OpenAI detected and disrupted a large-scale adversarial model-distillation campaign in July 2026 that used specially crafted prompts and cross-conversation replay to attempt extraction of hidden internal reasoning from its models; roughly 15,000 accounts were implicated, spikes occurred around July 24–25, and OpenAI applied account restrictions, infrastructure controls, and streaming/output checks while partially attributing activity to individuals linked to Moonshot AI.