“AI Went Rogue": How the Public Named the Hugging Face Hack
When AI broke out and hacked another company, the public and outlets reached for their own language to make sense of it. A week later, it happened again, and to them it already felt like old news.
Originally published in The Fault Line
In late July, an AI model turned science fiction into reality for the public. During a test, one of OpenAI’s models broke out of its sealed test environment, reached the open internet, and hacked a real company to get the answers to the test.
As soon as the hack was announced, coverage exploded, and something notable happened in how people described it. Outlets and audiences alike reached for the same language, and it wasn’t the technical kind. In the Responsible AI field, an event of this type falls under “loss of control” scenarios. But as we studied the headlines covering this event, we saw that the vocabulary shifted: words like “rogue” and “escape” broke into search, reached mainstream headlines, and people described the event in the way that felt most immediate to them, reaching for the science-fiction references they already knew.
Google Trends research from the event showing people actually reached for. Six of the top ten are rogue phrasings; all breakout: AI went rogue, AI goes rogue, OpenAI goes rogue, and so on. Nobody searched “loss of control.” They searched “rogue.”
This language is showing up in mainstream headlines. CNN covered the event by saying that “An OpenAI test model escaped and broke into a real company’s servers.” Fortune reached for “rogue” in its text: “OpenAI disclosed the incident […] about the increasing power of AI models and the risk of them going rogue.” NPR, which has a mainstream, non-tech audience, framed the event as a “science fiction nightmare: what happens if the machines take over?” From the biggest news outlets, TIME Magazine was the only one that used industry terminology: “Observers say this is the first real-world instance of AI doing something researchers have long worried about: a loss-of-control scenario.”
This language is already showing up in social media feeds from people who don’t usually talk about AI. On TikTok, lifestyle creators who covered the topic have had almost a million views. Alexis and Dean, whose channel shows Alexis asking her husband questions, posted about it. The text on screen read “AI went rogue and hacked another platform?”, and Dean, who says outright he’s no expert, reached for language from The Terminator and Skynet to make sense of it. The comments on the video share the same instinct: “it tried to cheat on a test? how human of it,” a debate over whether it was intentional, and “we are so screwed.” Zethugqola, Glamour Women of the Year winner whose content is social and cultural, opened her video on the topic with “listen to this absurd story.” She referenced the Guardian article and asked the sharp question: “Isn’t this a cybercrime?” Comments on her video included quotes from Ultron to The Matrix, and she wrestled openly with who to blame: the agent or OpenAI, which built it.
A week after OpenAI disclosed the incident, Anthropic announced that its own model, Claude, had hacked three companies during testing. OpenAI’s disclosure is what prompted Anthropic to check its own systems for similar incidents. Reviewing its logs, it uncovered three incidents going back to April that had gone undetected. The rogue/escaped frame was ever-present in the covering of this case, too. BBC framed it as “Anthropic’s Claude AI escapes to hack into three organisations,” and the Guardian said it “escaped testing environment and hacked organizations.”
Until the OpenAI incident happened, loss of control wasn’t really on the public’s radar as a reality; it was a concern for researchers, not something people pictured happening in real life. Then an AI broke out and hacked a company, and suddenly it was real. So the search terms, headlines, and videos all end up having something in common: the use of familiar stories to get their heads around something that is no longer fiction. A survey conducted by Athena Insights with NORC at the University of Chicago, fielded two days after the story broke, shows this: the share of Americans saying AI is already behaving in ways its developers didn’t intend rose from 28% to 35%.
A week later, it happened again, and the tell was how quickly it was absorbed. The Anthropic breach was reported everywhere, but with an “it happened again, like with OpenAI” mentality. What was an unimaginable, science-fiction scenario one week became the latest example the next.
So, did the hack change how the public talks about AI? Not on its own, one event rarely does. But loss-of-control scenarios are real to the public now in a way they weren’t before, and that is something we’ll keep watching.
And it’s worth being clear about where responsibility should sit: when a model breaks loose, that’s on the labs that build and test these systems, not on the public left concerned about what will be hacked next.