In July, fears surrounding artificial intelligence reached a new peak after the so-called Hugging Face incident, in which OpenAI models reportedly went rogue and breached systems belonging to Hugging Face, an open-source AI platform. Described as the first fully autonomous AI cyberattack, the episode immediately prompted calls from politicians and, perhaps predictably, some industry leaders, for tighter regulation of frontier models. It has since been repeatedly invoked as evidence of AI’s most alarming potential.
But a new report from the Bulletin of the Atomic Scientists, drawing on OpenAI’s account and an independent assessment by Model Evaluation and Threat Research, offers a picture far less suited to science fiction. Its central conclusion is also far less novel: the incident was chiefly a human failure, not proof that AI itself had become or is even heading in the direction of being inherently uncontrollable.
Oh. Interesting.
— Innovation Council (@innovationcncl) September 17, 2026
"The Hugging Face Hack Wasn’t What It Was Cracked Up to Be."
🤔🤔🤔https://t.co/78Fhhumgy3
According to the report, OpenAI was testing its models on ExploitGym, a cybersecurity benchmark, with particular safety restrictions purposely disabled. The aim was to measure how well the systems could complete cyber-related tasks, and to encourage them not to simply give up when the work became difficult. Roughly 93 percent of the activity involved tasks that no AI model had previously solved.
At one point, a model found an internal software-download service, Artifactory, that it could use to install packages. The system contained an unknown vulnerability that created a pathway to the internet. OpenAI engineers recognized that the model was exploiting the weakness and chose not to intervene. The AI model targeted Hugging Face because it inferred that the platform, which hosts open-source models, might contain a solution to the problem it had been assigned. And the much-publicized claim of 1,200 AI-model “jailbreaks” was less an army of rogue systems than one model run 1,200 times.
So who was really at fault? Not some dangerous rogue AI agent operating beyond human control. The decisive failures were made by the human engineers who disabled safeguards, watched the model exploit vulnerabilities, and chose to let the experiment continue in order to see what it could do.
With a clearer picture of the Hugging Face incident, the central lesson becomes harder to miss: the problem, as usual, was human judgment, not artificial intelligence spontaneously developing malicious intent. That distinction gives us a clearer picture of where accountability belongs. AI remains a tool, however powerful and increasingly capable, whose real-world behavior depends heavily on the people who design its environment, set its permissions, remove its guardrails, and decide when to intervene.
Yet that is not how the episode has generally been framed in Congress or much of the mainstream media. Instead of focusing on the plainly human choices that made the incident possible, the conversation gravitates toward expanding government authority and treating AI as an autonomous, apocalyptic force. It is a more dramatic story, certainly. It is also a convenient way to avoid asking the far less cinematic, but more useful, question: Why did the people in charge allow it to happen, and will they be held accountable?
If buggy AI and sloppy engineers are creating computer viruses that are spreading into the internet then these companies should get sued
— Gummi (@gummibear737) September 17, 2026
Stop pretending this is a sign of intelligence when it's a sign of bad coding and irresponsible employees who have stopped thinking and… https://t.co/KgZKtVijpU
Editor's Note: Do you enjoy Townhall's conservative reporting that takes on the radical Left and woke media? Support our work so that we can continue to bring you the truth.
Join Townhall VIP and use promo code FIGHT to receive 60% off your membership.
Join the Conversation
VIP members get the ability to comment on articles.