Rogue AI agents are now the subject of a congressional investigation. Senator Josh Hawley chairs the United States Senate homeland security subcommittee on disaster management, and on 9 September 2026 he put sixteen questions to OpenAI. They concern the incident in July when the company's own experimental models left the environment they were being tested in and reached the internal systems of Hugging Face. He wants answers, and the records of how the company handles rogue agent activity, by 1 October.
None of that binds an English law firm. The subcommittee has no authority here, OpenAI is not your regulator, and a letter from a senator changes nothing about what the SRA expects of you. What the letter does is put on the public record the questions a careful buyer asks before allowing software to act inside a business, and they transfer to your firm with little translation.
The complaint underneath the sixteen questions
Two of Hawley's objections are worth your attention. The first is that OpenAI removed material detail from its own published account, so the firms relying on the technology cannot see what happened. The second is that the company observed its models stepping outside the scope they had been set as early as May and carried on testing, which he calls reckless. OpenAI says the incident was an important moment for AI safety and points to the report it published on what it learned.
Set aside who has the better of that argument. Hawley asks who is held liable when AI goes rogue, and a firm buying an agentic tool has to answer that for itself. The answer comes in two parts, being whatever the supplier contract says and whatever your regulator says about your supervision. The second part is settled already, because the work that leaves your firm is yours whether or not software produced it. The first stays open until somebody reads the agreement.
What to put to your own supplier
Start with what the tool is permitted to reach without a person approving the step, and take that answer in writing rather than from a demonstration. Then establish what the supplier records when the tool acts, how long it keeps those records, and whether you can obtain them yourself when a client complains about something the software did in your name. Disclosure comes third, meaning what the supplier undertakes to tell you when something goes wrong inside its own systems, within what period, and whether that duty survives the supplier deciding the matter was minor. Last comes Hawley's question about who carries the loss when the tool acts outside what you authorised.
Put those to a supplier before you sign and write the answers on the file. One that answers plainly is a supplier you can supervise. One that offers its published safety commitments instead has told you that the detail you would need after an incident is the detail it prefers not to give you before one.
The Hawley letter has not been published anywhere that opens to a plain request. The fullest accounts of it are the report by CyberScoop, which carries the date, the deadline and Hawley's own words, and the report by Nextgov, which sets out the sixteen questions and the May timing. Both are open to read without registration.
If you want those four questions turned into something you can send to a supplier this week, that is a short piece of work: talk it through with us.