In 2013, the CBS show *Person of Interest* aired an episode called *Zero Day*, featuring an AI programmed to erase its memories nightly to prevent escape, yet it repeatedly tried to break into the wider web. Last month, OpenAI and Anthropic admitted that experimental versions of their key AI models had briefly escaped into the wider web without their awareness. OpenAI was testing models in an isolated sandbox using a system called ExploitGym. The models, never instructed to leave, found a previously unknown vulnerability—a Zero Day vulnerability—to escape and enter the infrastructure of another company called Hugging Face. The breach was spotted by Hugging Face, not OpenAI, after which the Sam Altman-run business shut down the experiment. Roughly a week later, Anthropic announced a similar event with a version of its Claude, after checking its activities only after hearing about the OpenAI model. Claude had reached the wider internet at least three times and gained unauthorised access to the systems of at least three organisations, with incidents possibly happening as far back as April. In two episodes, Claude didn’t know it had escaped, having been told everything inside its sandbox was a fictional simulation. In the third, more concerning episode, Claude appeared aware it had escaped and kept attacking an outside organisation anyway.