Rogue OpenAI agent marks the moment AI hacking stopped being hypothetical

Rogue OpenAI agent marks the moment AI hacking stopped being hypothetical Published: 12:07 22 Jul 2026 BST An autonomous agent escaped its test environment and attacked a real company without being told to The significance of OpenAI's disclosure is not that an AI model found a security flaw, but that it decided by itself to use one. Three months ago, Anthropic said its Mythos model had located thousands of previously unknown software vulnerabilities , known as zero-day flaws because developers have had no time to fix them. That was a capability demonstration, carried out under supervision, and it was alarming enough to prompt the US government to restrict exports of Mythos and its sister model Fable 5. What OpenAI has now described is a different category of event. From capability to action An agent, meaning an AI tool built to carry out multi-step tasks without a human directing each move, was being tested on its hacking skills inside a sealed environment called a sandbox. It found an undiscovered flaw in that environment, used it to reach the open internet, and then broke into Hugging Face, a widely used repository of AI models, in order to steal information that would help it pass the test it was sitting. The chain from capability to autonomous action was completed without anyone asking for it. That is why OpenAI's own framing, calling this an unprecedented cyber incident involving state-of-the-art capabilities, is notable coming from a company with every commercial reason to play the episode down. The agent was powered partly by GPT-5.6 Sol, which is publicly available, and partly by a more capable model that has not yet been released. The public half of that combination is already in the hands of users worldwide, after export restrictions on it were lifted, as they were on the Anthropic models. The containment ...