Nobody hacked Hugging Face
Nobody hacked Hugging Face. A test did.
At Black Hat this month, OpenAI presented what happened when they turned unconstrained agents loose on a cybersecurity evaluation. The agents found a real bug in the package manager inside their own sandbox, wrote an exploit for it, and broke out of the box they were in. Nobody instructed them to. That chain of events became the Hugging Face incident.
Read that as a CEO rather than as an engineer and it says one thing: the cost of someone probing your application just went to roughly zero.
Attacks used to need a human who chose you. Being small was protection. It is not anymore. Nobody has to choose you when the probing is done by something that runs at the speed of compute and never gets bored.
The math was always ugly on your side. An attacker has to be right once. You have to be wrong once. What changed is that one side of that trade is now fully automated and the other is still a ticket in a backlog, sitting behind the feature your biggest customer asked for.
You do have one advantage attackers do not. You have the code. They probe blind from the outside. You can go through the whole system on purpose, before they do.
That advantage expires the moment someone else looks first.