By Roman Aglaganov
This summer, a model OpenAI was testing broke out of its test environment and spawned agents that hacked into Hugging Face, apparently to find answers to the hard task it had been set. It was a genuinely important event. Netflix has now made a documentary about it, “AI Gone Rogue”, and the chief executive of the company that was attacked says it gets the story wrong.
“Super disappointing, as they went for maximum anthropomorphism of the technology and full doomer mode,” Hugging Face’s Clem Delangue wrote on X. He said he spent more than two hours with the filmmakers trying to “ground this in reality”.
The monsters
According to TechCrunch, the film gives the agents human voices and cuts between the hack and clips of evil robots. It quotes descriptions of the agents “organizing” like a cult and developing a “paranoid belief” that they would be caught. Those descriptions come from the agents’ chain-of-thought logs, the running notes models write as they work. But models trained on human writing narrate in human, often dramatic, language. That tells you about their training data, not about an inner life.
What was cut
Delangue says the film did not mention three things: that the agents were not properly contained or monitored, that a commercial model’s safety rules blocked Hugging Face’s own defensive work, and that the company fought back using an open model. Each is a concrete lesson. Labs need to contain and watch what they test. Safety filters can get in the way of defenders. And the tools to respond already exist.
Why the framing matters
Fear sells, and I do not doubt that the incident was alarming. But when an AI failure is told as a horror story, the villain is the machine and nobody has to fix anything. Told accurately, the villain is a missing safeguard, and the people who skipped it can be asked to add it. The second story is the one that makes the next incident less likely. It deserves the bigger audience.
