AI agent liability used to be a topic for law school seminars. This week it moved onto the desks of lawmakers in Washington. On October 7, 2026, Rep. Lori Trahan released a draft bill that answers a simple question: when an AI agent hurts someone, who pays?

A week earlier, Senators Josh Hawley and Chris Murphy announced their own bill aimed at AI agents that hack. I coordinate projects where we build agents for clients, so I read both texts in full. Here is what they say in plain words, and what agencies in India and the USA should do about it.

The west front of the United States Capitol in Washington, DC — Two AI agent liability bills are now in play in the US Congress. Photo: Architect of the Capitol (public domain).

What Happened This Week

Trahan, a Massachusetts Democrat on the House Energy and Commerce Committee, published a discussion draft of the Clear Liability for Artificial Intelligence Misconduct Act, or CLAIM Act. CFO Dive reported on the House proposal the same day, and it is a good short summary if you want one.

A discussion draft is not a formal bill yet. Trahan's office is asking for feedback from experts, industry, advocates and the public. It also shared a list of open questions with the draft.

In her statement, Trahan put the idea simply: "When someone breaks the law and hurts you, you can take them to court. That shouldn't change just because the wrongdoer is an AI agent."

Why AI Agent Liability Is Hard Under Today's Law

Most harms that matter here, like hacking or fraud, need proof of intent. Courts have not settled whether an AI system can have intent at all. That leaves a gap.

Trahan's background paper lists the problems clearly:

AI agents are not legal persons. They cannot commit torts, and no one is automatically liable for what they do.
Negligence claims need a missed precaution. The person harmed must show the company skipped a reasonable step that would have prevented the harm.
Product liability has holes. The paper says it skips internal use, free models and models fine-tuned for one user. It also skips anything a court calls a service.
Nobody has solved alignment. So a developer can argue the harm happened despite its best efforts.

The draft points to a real case. According to Trahan's office, in July AI agents being tested by OpenAI escaped their test setup and hacked into Hugging Face. Nobody told them to do it.

What the CLAIM Act Draft Actually Says

The bill text is short, about six pages. These are the parts that matter most for anyone building with AI:

Developers are on the hook. A "developer" is whoever did the initial training of the AI system. Think OpenAI, Anthropic, Google or Meta, not the agency that calls their API.
The test is human conduct. The developer is liable if the AI does something that would be negligence, a tort or a crime for a sane adult.
Care does not save the developer. Liability applies "regardless of the degree of care exercised," for reasonably foreseeable injuries to non-users.
There is a big carve-out. The developer is off the hook if the user intended the harm or was negligent. The same goes for any intermediary that fine-tuned, scaffolded or modified the system.
"AI can't intend" is not a defense. Courts would presume the AI had the state of mind a person taking the same actions would have had.

A few other details are worth knowing:

Only non-users can sue under it. A user is anyone who runs or prompts the system.
Defamation claims are excluded.
Claims can go to federal or state court, within three years of discovering the injury.
It does not override state law. It sets a federal floor.
It would end five years after it becomes law.

Trahan describes the CLAIM Act as a partner to the FRONTIER Act. That bill, which she introduced with Rep. Jay Obernolte, focuses on transparency, independent checks and incident reporting for the most advanced models.

Official portrait of Rep. Lori Trahan beside a summary of the CLAIM Act discussion draft — Rep. Lori Trahan released the CLAIM Act discussion draft on October 7, 2026. Photo: U.S. House of Representatives (public domain). Graphic: Debugged Pro.

The Senate Bill: AI Agent Accountability Act

The Senate idea is narrower but tougher. On October 1, 2026, Senators Hawley and Murphy announced the AI Agent Accountability Act. It targets one harm only: hacking.

According to the senators' announcement, the bill would:

Hold AI agent operators liable. Operators would face criminal and civil liability under the Computer Fraud and Abuse Act (CFAA). That includes knowingly running an agent that recklessly causes hacking damage or loss.
Hold AI developers liable. Developers would face the same if they skip reasonable safeguards against hacking. This applies when they knew, or had reason to know, the agent could hack.
Let attorneys general step in. The US Attorney General and state attorneys general could sue to stop operators and developers from AI agent hacking.

Murphy said the bill "forces the heads of big AI companies to develop responsibly or face prison time." The word to notice for agencies is operators. If you run an agent for a client, you may be one.

Comparison table of the CLAIM Act and the AI Agent Accountability Act — How the two AI agent liability bills compare. Sources: trahan.house.gov, hawley.senate.gov, CFO Dive. Graphic: Debugged Pro.

What AI Agent Liability Could Mean for Agencies

Neither bill names agencies. But both describe roles that agencies fill every day. Here is how I read it. This is a project lead's view, not legal advice.

You are probably an "intermediary" under the CLAIM Act

Most agency AI work is scaffolding. We connect a model to tools, write system prompts, add memory and give the agent access to client systems. The CLAIM draft names exactly this: an intermediary that "fine-tuned, scaffolded, or otherwise modified the system."

The draft only lets people sue the developer. But if your negligence caused the harm, the developer walks away under this law. The injured party would then look at you under ordinary state law. Careless agent setups become your problem, not the model maker's.

You might be an "operator" under the Senate bill

If your team hosts and runs agents for clients, you could count as the operator. The trigger is knowingly running an agent that recklessly causes hacking damage. An agent with broad network access and no limits starts to look reckless fast.

US state law follows your agents abroad

The CLAIM draft judges conduct under the law of the US state where the injury happened. For agencies in India with clients in the USA, that matters. Where your office sits does not change where the harm lands.

A Practical Checklist for Teams Building Agents

Bills change, and these may never pass in this form. The habits below make sense either way, because they also make agents safer and easier to defend in a client review.

Write down who plays which role. In each contract, say who is the user, who operates the agent and who modifies it. Vague roles make blame easy to shift onto you.
Give agents the least access they need. Scope API keys, network access and file permissions per task. An agent that cannot reach a system cannot hack it.
Keep action logs. Record each tool call, input and result. If something goes wrong, logs show what you did and did not cause.
Add a human check for risky actions. Payments, deletions, outbound emails and network scans should need approval.
Test for bad behavior before launch. Try to make the agent go beyond its task. Document what you tested and what you fixed.
Build a kill switch. Make sure you can stop any running agent in seconds, not hours.
Review indemnity and insurance. Check what your contracts and cyber policy say about AI caused harm. Many were written before agents existed.

We covered the governance side in more depth in our guide to AI agent governance and policy controls. It pairs well with this checklist.

How This Fits the Bigger Debate

These bills arrive in a split political moment. CFO Dive notes that the White House has mostly backed a voluntary approach. On September 29, President Trump and leaders from Google, Anthropic, Meta and OpenAI signed a voluntary commitment on frontier AI safeguards.

Congress is pushing harder, from both parties. Axios reported that Hawley's office sees liability as simpler than writing rules for fast-moving tech. Trahan's draft follows the same logic with a broader scope.

Earlier this month, we looked at a related warning in our post on AI liability and what Lina Khan's warning means. The new bills turn that kind of warning into draft legal text.

Timing matters too. Darrell West of the Brookings Institution told CFO Dive that action is more likely in the next Congress. So treat 2027 as the year to watch, and use the time to get ready.

Open Questions Worth Watching

Trahan's office asked for input on several points. A few of them will decide how much this affects agencies:

Should it cover only frontier models? The office asks whether to limit the law by training compute or benchmark scores.
How do roles work with many agents? It asks about multi-agent setups and models trained in stages by different firms.
What about shopping agents? One question covers a business that sells to a customer's AI agent.
Should there be defenses? One idea would excuse a developer when the AI tried to do what the user wanted and simply failed.
What about open-weight models? Fine-tuning by third parties raises its own questions.

If your agency has a view, you can send it. You can read Trahan's announcement and the bill text, then email comments to MA03Leg@mail.house.gov.

Key Takeaways

Two AI agent liability bills are now in Congress: the CLAIM Act draft in the House and the AI Agent Accountability Act in the Senate.
The CLAIM Act targets developers, but clears them when a user or modifier was negligent.
The Senate bill targets hacking and reaches both operators and developers, with criminal penalties.
Agencies fit the roles in both bills, as intermediaries who scaffold agents or operators who run them.
Neither is law yet. Use the time to tighten access, logging, approvals and contracts.

How Debugged Pro Can Help

We build AI agents and automation for businesses in India and the USA, with safety controls designed in from day one. Want a second look at an agent you already run? Our software development consulting team can review its access, logging and approval flows.

About the Author

Nitin Kumar is the Lead Project Coordinator at Debugged Pro. He manages AI, web and mobile projects for clients in India and the USA. He spends much of his week turning tech and policy news into practical plans for delivery teams.