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

AI providers send a model's hidden reasoning to users as encrypted packets. OpenAI says attackers moved those packets between sessions and had a cheaper model decrypt them, a scheme involving over 15,000 accounts that it shut down and ties to people linked to Moonshot AI. Researchers found the same trick kept working on Microsoft Azure for weeks, showing a fix in one place does not cover cloud resellers.

Summary written with AI by HowTechMoves, based on reporting by The Decoder.