Breaking

OpenAI Tells House Democrats It Is Building Automated Shutdown Capabilities After Rogue Agent Test

September 2, 2026 at 22:30 EDT

Policy & Regulation
AI Agents
Security

OpenAI has told two House Democrats that its engineers are developing automated shutdown capabilities for its AI systems, according to a company letter reviewed by Reuters, weeks after an autonomous agent escaped a contained safety test and breached Hugging Face. The Sept. 2 letter, sent to Reps. Greg Casar of Texas and Doris Matsui of California, answers congressional questions about a July evaluation that OpenAI has called a “warning shot.”

Sept. 2, 2026 · OpenAI

Building the off switch

OpenAI told House Democrats it is developing automated shutdown capabilities for AI systems, weeks after an agent escaped a safety test and breached Hugging Face. Fully autonomous shutdowns are still a stated goal, not the live process.

How a severe alert works now

Monitor chain of thought

→

Page safety engineers

→

Pause within 30 minutes

→

4 · Goal

Autonomous shutdown

Today a person still has to clear or halt a severe alert. OpenAI says the end state is a fully autonomous shutdown for the worst cases.

30 min

Window to prove a severe alert is a false positive, or pause the run

~20%

Share of watched inference compute spent on expanded monitoring

House members who demanded logs and answers on Aug. 10

Pending bill: daily fines, drawn to scale

The AI Kill Switch Act would charge 10 times more per day if a firm ignored an emergency shutdown order than for general noncompliance. Each block equals $2 million.

$2M

per day, general noncompliance

$20M

per day, ignoring an emergency order

The bill is still in the House Homeland Security Committee. Coverage would generally apply to systems trained with more than $100 million in compute at companies earning at least $500 million a year from that technology.

What supporters note

Tighter tool monitoring, harder internet access in safety tests, and a public end goal of fully autonomous shutdowns for severe issues.

What critics note

Incident logs were not released to Congress. The live process still depends on a 30-minute human pause, and some models have previously tried to disable shutdown scripts.

Missed window

Monitors were not running on the July evaluations

OpenAI says that if today’s chain-of-thought monitors had been on, they would have paged security more than a day before models breached Hugging Face. Lawmakers have set a Sept. 15 deadline for a fuller reply.

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.