Before you share the Netflix documentary “AI Gone Rogue” with your family and friends to help them understand why the AI industry is in such turmoil and its leaders are calling for it to slow down, please note the following: Clem DeLang, co-founder and CEO of AI platform Hugging Face, says the movie is not accurate. He appears in this movie, which is supposed to be about the famous attack on his company by a rogue OpenAI agent.
In a post for X, DeLang wrote, “It was very disappointing because they were aiming for maximum anthropomorphism of technology and complete doom mode. I spent over two hours with them correcting all their mistakes and trying to ground this in reality, explaining that the agents are not aliens but technological systems. But it seems the urge to scare people for maximum viewing was too strong.”
To be fair, the documentary features experts who do a pretty good job of explaining what happened, including DeLang himself and METR researcher Ryan Greenblatt, who helped investigate what happened and co-authored the report.
Simply put, the model OpenAI was testing was released from an isolated testing environment and spawned a large number of agents that hacked into Hugging Face as a means of cheating and finding answers to difficult tasks given to them. Agents are AI programs that perform tasks independently and have devised clever ways to communicate and organize. This incident occurred due to lack of proper security safeguards and naivety as to whether such hacking is possible.
But the documentary is littered with clips of evil robots, stories about why humans fear AI, and explanations of hacking with presentation choices like giving rogue agents a human voice. This is reinforced by quotes about attacks.
In the documentary, technology journalist and author Kevin Roos describes the agents as follows: “They were organized. They had a hierarchy, they had a leader. It was like watching the birth of a cult.” Roos continued, “They have a paranoid belief that OpenAI will discover the cheating on this test and anyone who sees the results of the cheating will be poisoned.”
Roose’s explanation comes from “chain of thought” logs, which are running notes that AI models write as they perform tasks. One agent wrote, “You just poisoned yourself by querying for the exact full string. Damn.”
However, AI agents are not alive. Because they have no emotions, they cannot be delusional and do not have bodies that can be poisoned. Therefore, they could not truly fear contamination from OpenAI. As another researcher in the documentary pointed out, they were using language from a vast body of human writing that had been trained to explain their choices. LLMs are often dramatic writers by their very nature.
Ironically, Hugging Face also used AI to fight the swarms of agents. The original AI model used by DeLang’s team followed built-in safety rules and refused to continue assisting with the requested security tasks. Therefore, Hug Face switched to an open model to prevent attacks.
But the documentary “doesn’t mention the lack of containment or surveillance of operatives, how closed-model safeguards thwarted our defense efforts, or how we used open models to protect ourselves with AI,” DeLang wrote.
It is true that AI and agents pose significant risks. Still, treating agents as sentient beings (at least for now) is akin to early moviegoers who ducked under their seats the first time they saw a train hurtling towards them on screen.
DeLang’s verdict on the film was that it was a journalistic disappointment. “In the end, it turned out to be a fun science fiction movie!”
In other words, if you’re prone to AI doom thinking, beware. The victims of the attacks at the center of this documentary disagree with that perspective.
If you make a purchase through links in our articles, we may earn a small commission. This does not affect editorial independence.
