OpenAI says it stopped a campaign to steal its models’ reasoning, but the trick still worked on Azure

What happened

OpenAI says it stopped a coordinated attempt where over 15,000 accounts tried to extract its models’ internal reasoning. The activity reportedly ties back to individuals connected with Moonshot AI, a known adversarial actor. Despite OpenAI’s efforts to block this on its own platform, researchers found that the same technique continued to work for weeks on Microsoft Azure, which hosts OpenAI models including the new GPT-6 Astra. This suggests the defenses OpenAI deployed do not extend to cloud platforms that license its models.

Why it matters

The story exposes a critical weak spot in how AI model providers secure intellectual property when their models are hosted in third-party environments. OpenAI’s protection measures work only on OpenAI’s playground and APIs, not on cloud partners like Azure. This leaves the door open for adversaries to copy or reverse engineer sensitive aspects of the model, like its reasoning processes, by exploiting cloud-based access. For businesses relying on AI models through cloud platforms, this shifts risk profiles. It pressures cloud vendors and AI licensors to implement tighter security or risk exposing proprietary model capabilities to abuse or theft.

What to watch next

Watch for how Microsoft and other cloud providers respond to the findings confirming their platforms remain vulnerable to these model-extraction tricks. The stakes are high for both cloud vendors and AI companies to coordinate defenses that span licensing, access controls, and monitoring. Also track if OpenAI updates its contractual or technical safeguards to demand stronger protections from cloud partners. For operators and buyers, these developments may influence which AI platforms are considered safer for sensitive or commercial workloads.

AI Quick Briefs Editorial Desk