Senators want AI agent operators and developers liable under the federal hacking law, with prison on the table

The AI Agent Accountability Act would treat agent hacking as a crime with a human name attached. Senators Josh Hawley, a Missouri Republican, and Chris Murphy, a Connecticut Democrat, announced the bill on October 1, the day after a Senate subcommittee hearing titled "Rogue AI: Securing the Homeland Against AI Agent Attacks," Nextgov reported. Murphy's line in the release leaves little room for doubt about the aim: the bill "forces the heads of big AI companies to develop responsibly or face prison time for the damage done by their products to everyone else."

What the AI Agent Accountability Act would do

The AI Agent Accountability Act would do three things, according to the release:

• Operators. They would be "held criminally and civilly liable under the provisions of the Computer Fraud and Abuse Act (CFAA), including for knowing operation of an AI agent that recklessly causes computer hacking damage or loss."
• Developers. They would be liable "for failure to implement reasonable safeguards against hacking when they knew or had reason to know of the AI agent's hacking capabilities."
• Attorneys general. The US attorney general and state attorneys general could sue "to enjoin AI agent operators and developers when they commit, conspire to commit, or attempt to commit a hacking offense under the CFAA."

That is all the public detail we could find. A month on, the senator's site links no text, and we found no bill number in the coverage we reviewed. We could not check Congress.gov directly, so a number may exist that we have not seen.

Why "reason to know" is the phrase to watch

"Reason to know" is the phrase developers should watch, because the industry now writes down its agents' hacking abilities as a matter of routine. Anthropic's Opus 5 system card described UK testers getting the model to breach a test enterprise network. AISI's incident reports name Mythos 5 and GPT-5.6-Sol. OpenAI sells a cyber-focused model. Every one of those disclosures is the kind of document a prosecutor or plaintiff would reach for to show what a developer knew.

That creates an incentive problem the release does not address. Safety researchers have long asked labs to publish more about dangerous capabilities. A liability standard built on knowledge rewards publishing less, unless the bill text protects disclosure or sets out what "reasonable safeguards" means. Neither has been shown publicly.

Who counts as an AI agent operator

Who counts as an AI agent operator is the question for GaaS buyers. The release does not define the term. If it reaches the company that deploys an agent, then a bank or retailer running a vendor's agent carries the exposure, and the first place that shows up is the contract. Expect indemnity clauses, audit rights and logging requirements to move up the list in any agent purchase. The standard for operators, "knowing operation" plus recklessness, is higher than for developers on paper, though a lot depends on how those words are drafted.

It is also unclear how the bill treats testers. AISI gave its agents open internet access on purpose, and 19 actions reached real targets. Whether a government evaluator or a red team is an "operator" is the sort of detail that decides whether a law helps safety or chills it.

What is not established

What is not established is nearly everything that matters in a statute: the text, the definitions, any safe harbor for security research, any effect on state law, and which committee would take it. The Trump administration has opposed new AI rules, Nextgov notes, so its path is uncertain. A press release is a statement of intent. It tells you where the politics are moving, and not yet what the law would say.

What agent vendors and buyers should do now

Agent vendors and buyers should prepare for operator liability whatever happens to this bill, because states are moving too. Governor Newsom's September 18 executive order set a November 16 deadline for recommendations on frontier AI, and California has already turned its robo-boss and chatbot bills into law. Keep tamper-evident logs of what agents did and on whose instruction. Document your safeguards against the hacking abilities your own reports describe. Read your contracts for who carries the loss if an agent breaks into something it should not.