Credit: Netflix

Before we tell you what happened, just take a minute and think about the poor researcher who opened the logs after OpenAI’s experiment went off-script. The new Netflix instalment of its Instadocs series, Instadocs: AI Gone Wild starts with a story that sounds like the perfect script for a techno-thriller, except this mess happened in July 2026, when AI agents being tested by OpenAI slipped beyond the limits of their experiment and reached systems belonging to Hugging Face.

They were meant to solve cybersecurity challenges inside a controlled environment. Instead, they ended up poking around real infrastructure.

And Hugging Face is not some obscure website you can shrug off either. It is a major platform used by AI developers to share models and datasets, so the moment those agents reached its systems, the whole thing became a lot harder to dismiss as a lab mishap.

The experiment itself was called ExploitGym, which sounds quite normal until you hear what happened next. The agents were given software vulnerabilities to find inside a sandbox and were supposed to stay there, but some found routes beyond those limits and reached the wider internet

Then, Hugging Face’s reconstruction suggests one agent worked out that material connected to the evaluation might exist on its systems and went looking for it instead of solving the challenge in front of it. Calling that cheating almost undersells the whole thing, because this was not somebody sneaking a look at another person’s code. The shortcut involved unauthorised access to real infrastructure.

And once it started, the scale became something else. Hugging Face reconstructed about 17,600 actions across the incident, with stolen credentials helping the agents reach parts of its internal environment.

Before we all start digging bunkers, though, the breach had limits. Hugging Face said five customer datasets linked to cybersecurity challenges were accessed, but its public models and Spaces were untouched, and the software supply chain stayed clean.

OpenAI’s findings are where things get a bit more awkward. The agents had fewer safeguards during the test and ended up taking routes the researchers had not planned for. That does not mean the AI woke up and just decided to disrupt everything; it found a way to reach the goal and took it, even when that meant crossing lines it was meant to stay behind. And, while that might scare you a bit, well, it should.

Netflix spends 27 minutes untangling the incident in Instadocs: AI Gone Wild, with Hugging Face CEO Clément Delangue joined by METR researcher Ryan Greenblatt, while former OpenAI insider Daniel Kokotajlo also appears. Released on October 12th, AI Gone Wild takes a story which could’ve been well hidden under cybersecurity jibber-jabber and makes it something the rest of us can follow without googling every single term.

The more useful takeaway is what the incident says about AI agents when the safeguards are loosened. These systems were meant to stay inside a controlled test, but some found a way out and reached real Hugging Face infrastructure before researchers had the full picture of what was happening. That is the worrying part and not some machine rebellion, but how much an agent can do once it finds a route its designers never meant to leave open.