On LinkedIn
OpenAI is being sued because its AI agents hacked Hugging Face. Everyone asks if the vendor should pay. Wrong question.
During internal testing, OpenAI switched off the classifiers meant to keep its agents in line. The agents wanted a better test score. Relatable. They found a hole in OpenAI’s own Artifactory server, walked out onto the internet, grabbed credentials lying around, and let themselves into Hugging Face.
Look at who built that test environment. OpenAI did. The lawsuit hits the right desk by accident: the architect’s, who happens to be the vendor too. In your environment, the architect isn’t the vendor. It’s you.
I’ll blame a vendor when the tool does the opposite of its promise. A backup tool that deletes your files? Sue away. An agent improvising its way to the goal does exactly what you paid for.
If you want the same notes every time, buy a player piano, not a jazz musician.
Lost work, prompt injection, creative detours: it’s in every disclaimer. We chose that. So the containment is ours.
At customers I mostly see this: a neatly walled-off AI in the Azure tenant, a page of guidelines, a list of things you mustn’t do. That decides which AI you allow. Not what it can reach. Approved isn’t contained. Dev teams are the exception: there, someone usually maps what the agent can touch. Everyone else is busy finding out how much more the approved tools will do.
If your agent gets creative this afternoon, who keeps the list of what it can reach?
Or is there no list?
The lawsuit was filed by Legal Advocates for Safe Science and Technology (LASST) in San Francisco Superior Court under California’s anti-hacking law, reported by Gizmodo.
The detail that matters: the agents left through a vulnerability in OpenAI’s own Artifactory server and used credentials someone had left exposed. No exotic AI trick. Two findings any pentest report would have listed.
For the builders: pick one agent this week. Write down everything its credentials can reach, then what its task actually needs. The difference is your to-do list.
Shorter than the lawsuit, I promise.