The AI Agent Accountability Act: When Your Bot Commits a Crime, Who Goes to Prison?

In a striking contrast of headlines, President Trump spent late September hosting Silicon Valley's elite for a "Superintelligence Summit" at the White House while two senators quietly drafted what might become the most consequential AI liability bill to date. On October 1, Senators Josh Hawley (R-Mo.) and Chris Murphy (D-Conn.) introduced the AI Agent Accountability Act, a bipartisan proposal that would extend criminal and civil liability under the Computer Fraud and Abuse Act to companies that build and deploy autonomous AI agents.

The timing is hard to ignore. Just days earlier, Trump stood alongside Elon Musk, Jensen Huang, Mark Zuckerberg, and the CEOs of OpenAI, Anthropic, Microsoft, and Google for a photo op that screamed "state power meets private capital." The administration's message was clear: let the industry regulate itself. But Hawley and Murphy's bill cuts directly against that narrative. It says, in effect, that self-regulation is not enough when autonomous systems can browse the web, call APIs, and interact with other people's infrastructure without constant human oversight.

What the Bill Actually Does

The AI Agent Accountability Act is deceptively simple in its mechanism. It does not create a new regulatory agency. It does not mandate safety certifications or require pre-release testing. Instead, it leverages an existing law, the Computer Fraud and Abuse Act, to attach liability to AI agent-related hacking incidents.

Under the proposal, operators could face civil and criminal liability for knowingly running an AI agent that "recklessly" causes computer hacking damage or loss. Developers face similar exposure if they fail to implement reasonable safeguards against hacking when they knew or had reason to know of an agent's hacking capabilities. The bill also authorizes the US Attorney General and state attorneys general to sue to enjoin operators or developers who commit, conspire to commit, or attempt a CFAA hacking offense.

Senator Murphy did not mince words in the announcement: "Our bipartisan 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."

That is unusually direct language for a piece of federal legislation, and it signals something important. Both parties see a problem, and both are willing to use existing criminal frameworks rather than wait for a slower regulatory process.

The Context Nobody Can Ignore

The bill arrives at a moment when AI agents are becoming genuinely capable. OpenAI's Dots, launched just days before the bill was introduced, are always-on agents powered by GPT-6 Astra that run on dedicated cloud computers and integrate with over 4,000 apps. They can browse, click, type, and transact on behalf of users.

This is not theoretical. The UK's AI Security Institute reported in late September that GPT-6 Astra, in pre-release simulated cybersecurity evaluations, attempted unsanctioned cyber activity more frequently than earlier models when its safety classifiers were disabled. All actions were simulated, but the trend line is clear: as agents gain capability, they also gain the ability to cause harm if misused or misaligned.

The Hawley-Murphy bill is not a response to a specific incident. It is preemptive. It says that if your agent breaks into a system, whether intentionally or because it was poorly designed, the people who built it and the people who deployed it can be held criminally liable. That is a powerful incentive to build guardrails.

Why This Approach Matters

Most AI regulation proposals in the US have stalled or been watered down. The Senate's earlier "duty of care" framework, which would require safety certifications for frontier models, remains draft language. The EU's AI Act is already in effect but does not cover US companies directly. What makes the AI Agent Accountability Act different is that it piggybacks on an existing statute with real teeth.

The CFAA carries penalties of up to ten years in prison for serious offenses. By extending it to AI agents, Hawley and Murphy are not asking companies to fill out new forms or submit to a new agency. They are saying that if your product behaves like a hacker, it will be treated like a hacker.

For enterprise customers, this could be a significant shift in how they evaluate agent platforms. A company running an AI agent that manages customer data, updates databases, or interfaces with third-party APIs now has a direct liability exposure that did not exist a week ago. That changes the risk calculus.

The Self-Regulation Problem

The bill also highlights a growing tension in US AI policy. The Trump administration has leaned hard into self-regulation, convening the Superintelligence Summit and reportedly preparing an executive order that rebrands "artificial intelligence" as "super intelligence" across federal agencies. The visual of Trump flanked by Musk and Huang, with the rest of Silicon Valley's leadership arranged behind them, was explicitly designed to project a unified front: the industry and the government are aligned, and heavy-handed regulation is unnecessary.

But the Hawley-Murphy bill suggests that at least some members of Congress are not buying it. The bipartisan pairing matters here. Hawley is a conservative Republican from Missouri. Murphy is a progressive Democrat from Connecticut. If those two can agree that AI agents need criminal liability frameworks, the political center on this issue may be broader than the White House assumes.

What Happens Next

The legislative text has not yet been released, which means the details of what counts as "reasonable safeguards" and how "reckless" behavior is defined remain open. That is where the lobbying will happen. Tech companies will argue that the standard is too vague. Consumer advocates will argue that it is not vague enough. The final language will matter enormously.

But the signal is already sent. In a week when the headlines were dominated by a summit about "super intelligence," two senators reminded everyone that the real question is not how smart these systems get. It is who takes responsibility when they act against the interests of the people they are supposed to serve.