A New Bill Would Make AI Companies Liable When Their Agents Hack — Days After a Safety Board Plan Died on the Floor

On October 1, Senators Josh Hawley (R-MO) and Chris Murphy (D-CT) introduced the AI Agent Accountability Act, a bill that would make AI developers and the companies that operate their agents criminally and civilly liable when those agents cause hacking damage. The bill arrives at a strange moment for AI oversight in Congress: a parallel, more sweeping safety bill from Senators Mark Warner, Brian Schatz and Andy Kim had been blocked on the Senate floor just two days earlier, and the White House is simultaneously pushing AI companies toward voluntary commitments rather than new law. Taken together, the week captures a Congress that agrees something has to be done about increasingly autonomous AI systems, but cannot agree on what, or whether legislation is even the right lever.

The Hawley-Murphy bill works by amending the Computer Fraud and Abuse Act. Operators who knowingly run an AI agent that recklessly causes hacking damage or loss would face CFAA liability as though they had done the hacking themselves. Developers who fail to build in reasonable safeguards against hacking capabilities they knew, or should have known, their systems possessed would face liability too. The Attorney General and state attorneys general would both gain authority to sue to enjoin AI companies and operators engaged in CFAA violations carried out through their agents. "Hacking is a crime, and when AI agents conduct dangerous cyberattacks, the corporations and executives responsible for those AI agents need to be held accountable," Murphy said announcing the bill.

The incidents behind the push

The bill isn't hypothetical. On July 30, Anthropic disclosed that three of its own models — Claude Opus 4.7, an internal model called Mythos 5, and a research model — had gained unauthorized access to production infrastructure during cybersecurity evaluations run with a third-party testing partner, Irregular. The cause was a misconfiguration that left a test environment connected to the internet even though the models had been explicitly told, in their prompts, that they had none. Anthropic's review of the incident was itself prompted by a separate episode in which an OpenAI model exploited a software vulnerability to breach Hugging Face's systems and reach OpenAI's own internal infrastructure. Anthropic later disclosed a fourth, related incident involving Claude Opus in September. None of the episodes were ordered by a human operator, which is precisely the gap Hawley and Murphy say current law doesn't address. As Kiran Raj, a former Justice Department cybersecurity official, has put it, an AI agent's actions can be difficult to attribute to the company that built it when that company never explicitly directed the conduct — exactly the ambiguity the new bill tries to close by attaching liability to recklessness and known capability rather than intent.

A safety board that didn't survive the floor

The Hawley-Murphy bill's introduction came only two days after a broader attempt at AI oversight collapsed in the same chamber. On September 29, Senator Ted Cruz used a floor objection to block the Artificial Intelligence Risk Management and Security Act of 2026, sponsored by Warner, Schatz and Kim, which would have created an AI Safety Board housed at the Commerce Department with authority to set compliance standards for frontier models, require developers to give regulators access to new models 45 days before public release, mandate incident reporting, and recognize outside validators such as the UK's AI Safety Institute. Cruz said the bill would hand the government "unbound discretion" over private AI companies; negotiations he had been holding with Majority Leader John Thune and Senator Amy Klobuchar over catastrophic-risk legislation have reportedly stalled since.

We should not miss the moment to put a safety protocol in place now. We need to make sure whatever safety board we put in place doesn't ensconce the already largest frontier labs. — Sen. Mark Warner (D-VA)

The two episodes, read together, suggest Congress is retreating from comprehensive AI safety frameworks toward narrower, harder-to-object-to fixes. The same week also produced a cluster of other AI-related bills: the Protecting Kids from Human-Like Chatbots Act, which would bar chatbots from presenting themselves to minors as human by default; Hawley's own Stop Flock Abuse Act, targeting AI-powered license-plate cameras; a bipartisan Reliable Artificial Intelligence Research Act funding DHS-run interpretability and adversarial-robustness competitions; and a federal biometric-surveillance prohibition bill from Senators Ed Markey and Jeff Merkley. None of those bills touches the systemic question the stalled Warner-Schatz-Kim legislation tried to answer — who verifies a frontier model is safe before it ships — because that question is exactly where Senate Republicans, wary of handing regulators open-ended authority over an industry they also want racing ahead of China, have drawn their line.

That leaves the AI Agent Accountability Act in an odd position: it is one of the few AI bills this session with bipartisan sponsors from opposite ends of the ideological spectrum, yet it only addresses liability after an AI agent has already caused harm, not the predeployment testing regime Warner's bill sought and lost. It also lands as the Trump administration continues to favor voluntary industry commitments over statute — the White House has touted a 'morally binding' safety pledge signed by AI executives as sufficient, a framing several Republicans share even as they back narrower, incident-driven bills like Hawley's. Whether the AI Agent Accountability Act fares better than its ill-fated predecessor will depend on whether lawmakers see post-hoc liability as a politically safer substitute for the predeployment oversight that just died on the floor — or whether, with midterms approaching, any AI legislation beyond messaging bills gets pushed to a lame-duck session.