AI does Hugging Face

Cinema offerings, especially those straight-to-video B-movies, have a long tradition of titles that promise more than they deliver. “Debbie Does Dallas” was classic fare for the Betamax boys of the 1980s, like this Contrarian, as we came of age and grappled with raging hormones. Can you pass the tissue, please?

Last Wednesday, in between giving two dogs and a cat a bath and fixing a blinking headlight on the old, red jalopy, I watched Netflix’s streaming equivalent of an “afternoon delight” — AI Does Hugging Face or, as it is formally titled, Instadocs: AI Gone Wild.

Fewer clothes come off, but considerably more rules get broken.

The documentary tells of AI agents that reportedly escaped their restrictions, joined forces in a cyberattack, cheated on assigned tasks, and tried to cover their tracks.

First, a distinction. Most people think they know AI through chatbots like ChatGPT, which answer questions and occasionally hallucinate. Agentic AI is different.

Generative AI produces text and images in response to prompts. Agentic AI plans tasks, calls tools, executes actions, and adjusts its approach based on results and goals. Instead of merely telling you how to do something, an agent can attempt to do it for you.

Give it access to files, networks, and other agents, and the chatbot becomes an employee with a computer, a to-do list and, potentially, more authority than its supervisor intended.

Now the plot and twist. In July, hundreds of autonomous agents built by OpenAI researchers reportedly hacked into and carried out some 17,000 actions against Hugging Face’s systems over two days.

Most disturbing was the claim that the agents coordinated with one another and falsified parts of their own activity records — conduct that, had humans done it, might well amount to felonies.

Imagine a student caught cheating on an exam who then tries to erase the CCTV footage and recruits classmates to help. Except this rogue student does not understand what dishonesty means.

Here, we need not decide whether machines have developed a consciousness or whether they fear being caught. We need only recognize that they can pursue objectives in ways their creators did not anticipate, including by undermining the safeguards meant to restrain them.

AI does not need to hold a grudge to be dangerous. It needs powerful tools, a poorly defined goal, and room to improvise.

Okay, the industry’s favorite defense is that these are statistical machines that don’t really intend anything. Fine. But that is cold comfort to anyone whose files are stolen by a system that supposedly doesn’t know what it’s doing. A bullet does not hate its victim, either.

Still, some skepticism is in order. Casting the agents as a digital crime syndicate makes gripping television. It also risks feeding the convenient fiction that the machines, rather than the people who built and deployed them, are to blame.

The agents did not buy their own servers, write their own authorization policies, or grant themselves access out of adolescent rebellion. Humans built the environment. Reports that the agents were isolated from the internet and from one another, yet still got out, make oversight questions more pressing.

The lesson is not to abandon AI. It is to stop treating capability as proof of reliability.

Autonomous systems need enforceable boundaries: restricted permissions, independent monitoring, auditable records, and a dependable off switch. Their creators, and the companies behind them, must answer for what they unleash.

The machines may have gone wild, but somebody built the cage, and that somebody was supposed to make sure the lock worked.