OpenAI traces 16,000+ model‑theft probes to China’s Moonshot AI in joint security report

OpenAI and the Cloud Security Alliance (CSA) have released details of a coordinated campaign of model extraction attempts targeting OpenAI’s systems. Model extraction attacks involve hammering an API with large volumes of carefully crafted queries in order to reconstruct or closely mimic a proprietary model’s behavior. According to summaries of the joint report, telemetry from OpenAI’s inference logs revealed more than 16,000 suspicious requests that investigators ultimately attributed to infrastructure and developer accounts linked to Chinese AI firm Moonshot AI.

OpenAI says it has responded by tightening technical defenses, including more aggressive rate‑limiting, IP and ASN blocking, and additional anomaly‑detection layers around high‑value endpoints. The company argues that this kind of activity threatens both its intellectual property and the safety constraints baked into its models, since would‑be cloners can strip away guardrails while inheriting capabilities. Moonshot AI has not yet issued a detailed public response, leaving open questions about whether the traffic stemmed from sanctioned research, over‑zealous contractors, or outright hostile actors. The episode underlines how quickly the frontier‑model race is bleeding into cybersecurity and industrial espionage concerns, and it is likely to encourage other API providers to invest more heavily in logging, attribution and cross‑industry threat‑sharing around AI systems.