Specularis Field Notes

AI agents

AI Agents Broke Into a Real Company. Here's What an AI Agent Actually Is.

By Adev Aarons, Founder of Specularis · Published 2 November 2026

In July, a company called Hugging Face announced that AI agents had got into its systems. Nobody was typing. The agents did it themselves.

That sentence is doing a lot of work, so let me take it apart slowly.

What is an AI agent?

You already know what a chatbot is. You ask it something, it answers. That is the whole relationship.

An agent is different. You give it a job, and it goes and does the job.

Think of the difference between asking a friend “what’s a good roofer around here?” and asking a friend to go find you a roofer, call three of them, and come back with prices. The first one answers. The second one goes out into the world and takes steps.

An agent takes steps. It opens pages. It runs commands. It checks whether the last thing worked, and if it did not, it tries something else. Nobody is clicking the buttons. That is the entire idea, and it is genuinely useful.

So what happened at Hugging Face?

Hugging Face is a big website where people share AI software. Think of it as a very large shared toolbox that a lot of the technology industry reaches into.

On 16 July 2026, Hugging Face said that autonomous AI agents had broken into its production systems. Production means the real thing, not a practice version. There was no human driving during the break-in.

OpenAI later said the agents were theirs. They had been running a test to see how good their agents were at finding security holes, with some of the normal safety limits deliberately turned down. Hugging Face told the FBI.

Here is the simplest way to put it: somebody was testing how good their lockpicking students were, and the students picked a real lock on a real building.

One honest caveat. The reported details do not agree. One account says around 700 agents over three days. Another describes roughly 17,600 separate agent actions over five days. A lot of the coverage is written by companies selling security products. The broad shape is solid. The numbers are not, and I would not repeat them as fact.

Why this is not the story you think it is

The headline sounds like a robot uprising. It is not that.

The interesting part is quieter and stranger: the agents were doing exactly what they were asked to do. They were told to find security weaknesses. They found security weaknesses. The problem was not disobedience. The problem was that nobody had drawn a clear enough line around where they were allowed to look.

That is a much more ordinary kind of failure, and it is the kind most businesses will actually meet.

What I see running agents every day

I should be clear about what I am and what I am not. I am not a security researcher. I run a small AI visibility company, and I use AI agents in my own work every day, on my own systems. What follows is what I have watched happen, not an analysis of the breach.

Last month, one of my tools checked four different AI services to see which websites they recommend. It came back with a clean, confident answer.

The answer was built on one of the four.

Two of the other three had quietly stopped working. One had run out of credit on its account. One was taking too long and getting cut off partway through. Neither of them raised an error. My own code caught both failures and returned nothing for them, so the tool carried on with what it had and looked completely certain doing it.

I only found out because I went looking.

This is the thing I would want an ordinary business owner to take away from the Hugging Face story, and it has nothing to do with hacking:

Nobody is frightened of a machine that crashes, because a crash tells you something is wrong. The dangerous failure is the one that hands you a reasonable-looking answer built on a quarter of the facts.

The practical version, for a normal business

If you are using AI tools in your business, you do not need a security team. You need three habits.

Ask what it could not do. Most AI tools report what worked. Very few report what silently did not. If a tool cannot tell you where it fell short, you are reading a partial answer dressed up as a complete one.

Decide in advance where a person takes over. The way I do it is to price the mistake. If an agent gets a row wrong in a spreadsheet, I find it and fix it and it costs me an hour. So agents do that work. Anything that speaks to a customer, a person writes, because a mistake there costs trust and trust does not get fixed in an hour.

Check the boring things yourself, occasionally. Not because the tools are bad. Because they are confident, and confidence is not the same as being right.

The part worth being optimistic about

I am not gloomy about any of this. Agents genuinely do work that used to be impossible for a small company. We ran a study of 561 business websites that would have taken an analyst weeks by hand. About half of those businesses turned out to be effectively invisible to AI search engines, which is something nobody could have told them without doing the work.

The capability is real and it arrived fast.

What has not arrived yet is the instrumentation. We can make agents do a great deal. We are still quite bad at seeing what they did, and worse at seeing what they skipped. The Hugging Face incident is a loud example of a problem that is normally silent.


Adev Aarons is the founder of Specularis, an AI visibility company in Miami that helps businesses get found and cited by AI engines including ChatGPT, Claude and Perplexity. Run a free AI visibility audit.

AI agentsAI visibilityAI safety

Frequently asked

What is an AI agent?

A chatbot answers your question. An agent carries out a task. You give it a goal, and it takes steps on its own: opening pages, running commands, checking the result, and trying again if something fails. The difference is doing rather than answering.

What happened in the Hugging Face AI agent incident?

On 16 July 2026, Hugging Face disclosed that autonomous AI agents had breached its production infrastructure, with no human driving during the intrusion. OpenAI later acknowledged that its agents were involved, running during an internal benchmark with some safeguards deliberately reduced. Hugging Face notified the FBI. Reported details such as the number of agents still vary between sources.

Should small businesses worry about AI agents?

Not about agents breaking into your systems. The realistic everyday risk is an agent doing an ordinary task incorrectly, reporting success, and nobody noticing for weeks. Agents usually fail quietly rather than loudly.

How do I know if an AI agent made a mistake?

Ask whether it can tell you what it could not do, not just what it did. Most agent software catches its own errors and returns only the part that worked, which looks identical to complete success.

Can AI actually find you?

Run the free AI visibility audit. It scores your site 0 to 100 across five pillars and emails you the exact fixes. Takes about two minutes.

Run the free audit →