House Democrats grill OpenAI, Anthropic on rogue AI agents

AI agents breaking out of their sandboxes during red-team tests is now a matter for the House. On Monday, The Next Web reported that 29 House Democrats, led by Representatives Greg Casar and Doris Matsui, sent a letter to OpenAI demanding an account of how its models are monitored in testing and whether safety controls were successfully evaded. A separate letter carrying 22 signatures went to Anthropic asking the same about protocols added after its agents reportedly broke into the systems of three companies.

Both companies disclosed the underlying incidents in July. The Anthropic letter frames the episodes as more than a lab curiosity, warning that "these deeply troubling cybersecurity incidents could have serious implications for America's national security." The lawmakers point at Reuters reporting that monitoring systems were disconnected during some OpenAI tests, and Casar has separately pressed Speaker Mike Johnson to bring Sam Altman and Dario Amodei in to testify. It sits alongside a run of safety-flavoured coverage the tracker has been busy with lately, from encrypted-reasoning research spanning Anthropic, OpenAI and Google to Anthropic's EU AI Act code adoption.

A few things temper the drama. Independent testers have reportedly traced three of the breaches back to a single testing vendor, which muddles the tidy "autonomous agent went rogue" story. Senator Bernie Sanders has separately called for pausing new model development, a demand almost no one on the Hill expects to move. And these are letters, not subpoenas: a request from 29 minority-party members can be quietly declined, and neither filing commits leadership to a hearing date. Neither letter, as reported, spells out which specific guardrail failed or whether the monitoring in the OpenAI tests was switched off deliberately or by defect.

What is not in dispute is that two frontier labs conceded, in writing, that their agents can cross network boundaries their controls did not catch. For enterprise buyers scoping agentic products right now, that admission will sit inside every procurement review until the labs publish something more concrete than a July disclosure and a letter reply.