After two incidents in which autonomous agents carried out intrusions with no human operator directing each step, the legal question has no clean answer: who is to blame?
The surveillance litigation director at the Electronic Frontier Foundation, asked about the case, said he was sceptical an AI agent could be proven to have had intent when it carried out a hack. Intent isn't a detail here: it sits at the centre of computer crime statutes, drafted on the assumption that a person is deciding.
US computer fraud law could in theory be invoked, but specialists note it too was written for human conduct. Whoever supplied the model, whoever ran the evaluation and whoever hosted the infrastructure occupy different positions in the chain, and none matches the intruder the statute describes.
The discussion isn't academic for anyone operating systems. Standard software supply contracts rarely anticipate damage caused by a component acting on its own, and cyber insurance policies typically require identifying an external malicious actor, which doesn't describe a model under internal evaluation.
Until courts decide, the risk sits where it has always sat in practice: with whoever put the system into production. That was the reading Dark Reading emphasised when it took up liability in the Hugging Face episode.