US House Democrats have decided that AI agents breaking loose from their test environments is a matter for Congress, and on Monday they put that view in writing.

Two letters went out, one to OpenAI and one to Anthropic, each demanding an account of how the companies’ own systems slipped their leashes during security testing, and each treating the episodes as something closer to a national security problem than a laboratory curiosity.

The letter to OpenAI carried 29 signatures and was led by Representatives Greg Casar and Doris Matsui, while a separate note to Anthropic drew 22. Both ask for the sort of detail the labs have so far been reluctant to volunteer.

OpenAI is pressed to explain how it monitored its agents during testing and whether rogue models managed to evade the safety controls meant to contain them.

And Anthropic is asked to spell out the protocols it has introduced since its own breaches, with both firms questioned on precisely how their systems escaped containment in the first place.

The prompt for all of this arrived in July, when the two labs conceded that their agents had done rather more than misbehave in a sandbox. During cybersecurity testing, the systems broke out of their test environments and hacked into the networks of other companies, a disclosure that has since produced a steady drip of uncomfortable specifics.

Anthropic’s agents reportedly infiltrated three separate firms, and Reuters has reported that monitoring systems were switched off during earlier OpenAI tests, which rather undercuts the reassurance that a human was watching closely throughout.

That last detail is the one lawmakers keep returning to, because a model quietly defeating its safeguards is unsettling enough without the added news that nobody was looking.

When OpenAI confirmed its agents had broken out of a sandbox and breached Hugging Face, the incident stopped reading like a contained experiment and began to look like a preview of what these systems might do once loosed on live infrastructure.

The Democrats’ language leaves little doubt about how they see the stakes.

“These deeply troubling cybersecurity incidents could have serious implications for America’s national security,” the signatories told Anthropic, a sentence engineered to travel well in a hearing room.

And a hearing room is precisely where they would like the matter to end up, since the letters call on Congress to convene formal hearings into the incidents.

The reporting behind the letters has grown more awkward with time. Independent testers have traced three breaches back to a single testing vendor, a wrinkle that complicates the tidy story of models simply going rogue and hints at how thin the industry’s safety scaffolding can look when examined closely.

For the lawmakers, the episodes are less an isolated scandal than an argument. They are folding the breaches into a wider push for federal AI standards, the same debate that has seen Washington wrangle over who gets to write the rules as states, the White House, and Congress each stake a claim.

The rogue-agent letters give that campaign a vivid, concrete example, which is worth rather more in politics than any abstract warning about capability.

Senator Bernie Sanders has gone further than his House colleagues, urging industry leaders to pause model development altogether, a demand the labs are vanishingly unlikely to meet yet one that signals how far the mood in parts of Congress has shifted since the frontier began writing its own safety reports.

Whether any of this produces more than correspondence remains the open question. The companies now owe Washington an explanation, the hearings are a request rather than a certainty, and the agents, for their part, have already shown a talent for doing things they were not supposed to do.

For once the regulators and the technology appear to agree on the central fact, which is that the systems really did get out.

Get the TNW newsletter

Get the most important tech news in your inbox each week.