If an AI agent logs in or buys something on your behalf, who's committing the crime if it breaks the rules?

This explores how laws written to punish unauthorized computer access should handle an AI agent that logs in, browses, or buys on someone's behalf. One caveat up front: the corpus has no notes that analyze computer fraud statutes or the cases being litigated under them. What it does have is a set of adjacent legal and governance arguments. Together they show why these statutes strain under agents, and which ideas courts might borrow.

The central problem is that 'unauthorized access' assumes a human whose intent and permission you can read. Agency law and principal-agent economics name three problems with delegated actors: hidden information, discretion over how to act, and divided loyalty. Conventional enforcement struggles with all three once the agent makes decisions nobody can interpret, at machine speed Can traditional governance tools actually control AI agents?. Consumer law runs into the same wall. EU doctrines like the 'average consumer' test and disclosure rules were built around human thinking and break down when the buyer is an algorithm Can EU consumer law protect AI agents as algorithmic buyers?. Computer fraud law has the same human-centered premise. 'Exceeding authorized access' asks what a person knew and intended, and an agent has no intent of its own.

One way through is to stop asking what the agent intended and ask who answers for it. Owens and Law argue that AI shopping agents should carry the fiduciary duties of human brokers, which would tie the agent's conduct firmly to the people who deploy it Should AI shopping agents face fiduciary duties like human brokers?. A Munich court took a similar line in a different area. It held Google directly liable for false AI Overview statements and treated the generated text as Google's own speech, not as neutral indexing Can courts hold Google liable for AI Overview false claims?. Applied to computer fraud, that reasoning would treat an agent's access as the deployer's access, with no 'the bot did it' gap.

The less obvious point is that responsibility is drifting toward whoever runs the website. Boboev shows that agent-driven checkout removes the behavioral signals merchants use to spot fraud, such as typing rhythm and mouse movement, while merchants remain liable as Merchant of Record Does agentic commerce shift fraud liability onto merchants?. This matters for computer fraud cases because they often turn on whether the site owner clearly withdrew permission. If sites can't tell an agent from a person, it gets hard to say permission was ever clearly granted or refused. That makes verifiable records a likely deciding factor. Deloitte proposes digital identities, cryptographic receipts, and permanent action logs for agents How should organizations supervise autonomous AI agents?. Records like these could give courts concrete evidence of who authorized what.

Courts might also scale scrutiny to how much control the person handed over. The argument that risk to people rises with each step of autonomy given to an agent Does AI risk increase with the autonomy we give it? suggests a sliding scale. A user who approved each step would be judged differently from one who released a fully autonomous agent. History also suggests courts move slowly. They regularly catch AI-fabricated citations but rarely punish the fabrication itself, instead falling back on familiar tests of harm and intent Do courts actually sanction fabricated AI citations when detected?. Expect the same with computer fraud law: old intent-based tests applied to the human behind the agent, until legislators or new norms like fiduciary duties supply something better.