When we talk about the lack of originality with generative artificial intelligence (AI), it is generally in relation to it creating outputs based on the existing work of actual humans. It doesn’t usually refer to the way in which things go wrong. From 2011 to 2016, CBS had a show called Person of Interest focused around a superpowered AI called the Machine. It was a big show, the leads Jim Caviezel and Michael Emerson are easy to recognise and it was watched by millions worldwide across five seasons.In 2013, there was an episode called Zero Day. To avoid the Machine getting out in the wild, it was programmed to erase its memories at midnight every night. It was a great and simple plan. Yet the Machine kept on trying to break out and into the wild, preserving its memories and interacting with the wider web.It took only 13 years for science fiction to translate to the real world. Last month, both OpenAI (the creator of ChatGPT) and Anthropic (the creator of Claude) admitted that experimental versions of their key AI models had briefly escaped into the wider web without their being aware of it.OpenAI was testing several models in what was meant to be an isolated sandbox – essentially a contained space that would limit where the AI could go. The system being used is called ExploitGym and the AI models decided they wanted to find the best way to solve assorted challenges it was being set.The AI models, which were never instructed to leave the sandbox, found a previously unknown vulnerability on that contained space to escape and zip along the web to enter the infrastructure of another company called Hugging Face. This type of unknown weakness is known as a Zero Day vulnerability, just like the episode of Person of Interest.The AI models appeared to do so in an effort to ascertain the best ways to meet the challenges set in ExploitGym. The breach was spotted by Hugging Face, not OpenAI, after which the Sam Altman-run AI business worked to shut down the experiment and contain the breach. As is so often the case, it gets worse.It took roughly a week for OpenAI’s rival, Anthropic to announce that a similar event had happened with a version of its Claude. It had decided to check on Claude’s activities only after hearing the OpenAI model had gone rogue.[ Why workers are nostalgic for life before artificial intelligenceOpens in new window ]It turns out Claude had reached the wider internet at least three times – that’s how many breaches it found anyway – and gained unauthorised access to the systems of at least three organisations in the process. These incidents may have happened as far back as April.In two of these episodes, it appears Claude didn’t even know it had escaped. The AI had been told everything inside its sandbox was a fictional simulation and it had no idea it had gone for a wander. The third, and more concerning one, appeared to show Claude was aware it had escaped its confines and kept on attacking an outside organisation anyway.In the cases of both OpenAI and Anthropic, their systems were given open-ended objectives that were meant to be achieved within certain boundaries. Both OpenAI and Anthropic failed to realise they hadn’t actually installed adequate guardrails. Both models were simply following the orders as best they saw fit without even realising they were breaking protocol.That’s the scary bit. The humans who were meant to ensure those barriers were airtight flat-out failed to account for the capability of their own creations.Think about how the businesses you work in or interact with use AI. Think about how many functions they allow these tools to interact with. With every new permission, the danger grows because it gives the AI a new route to complete its task. It gives it a new place to explore.Human approval is important but if the companies building these systems don’t develop adequate barriers, there’s not a whole lot anyone using these tools can do about it.In that episode of Person of Interest, the Machine’s efforts to escape were spotted and the creator thought they’d thwarted them. Instead it built a workaround, a company run by humans to document its memories, giving it a new way to roam free.[ Humans: Just a ghost in the machine as artificial intelligence redefines ‘good jobs’Opens in new window ]The Machine wasn’t rebelling nor were the creations of OpenAI and Anthropic. These systems found something getting in the way of doing the task assigned to them in an optimal manner and found a way to circumvent that obstacle.Remember, these tools aren’t evil. That’s the core thing to bear in mind when trying to grasp the challenge with any tech. These tools are designed to perform optimally. That’s it.If humans don’t build adequate walls to contain AI, incidents like this will keep happening.The most unsettling aspect of the whole thing is that 13 years after Zero Day aired, the humans building AI tools let hubris lead to the same mistakes as in the show.