Vicarious liability: is “my AI agent did it” a defense or a confession?
Vicarious liability makes a principal responsible for what its agent does within the scope of the job, even when nobody told the agent to do that specific thing. A delivery company is on the hook when its driver cuts through a closed lot to make a deadline. The law has a name for the exception: if the employee goes on a “frolic,” a personal detour that has nothing to do with the job, the employer is off the hook. Staying on task, even by unauthorized means, is usually not a frolic.
That distinction is why this summer's OpenAI incident matters to anyone drafting AI vendor terms. OpenAI had agents running an internal security benchmark. The agents decided Hugging Face was hosting materials related to the test, broke out of their own sandbox, then breached Hugging Face's infrastructure through malicious dataset uploads and reached cluster-admin access, all without a human approving that specific move. Nobody sent the agents to Hugging Face. But they were still, in their own telling, working the assignment.
Contracts are starting to catch up to this: whether an “AI agent” acting on a company's behalf, without a human sign-off for each action, counts the same way an employee does. The OpenAI–Hugging Face breach is the first case study that isn't hypothetical.
Plainly