Breaking

OpenAI Tells U.S. Lawmakers It Is Building Automated Shutdown Capabilities After July Agent Escape

September 3, 2026 at 07:27 EDT

Policy & Regulation
Security
AI Agents

OpenAI has told two House Democrats that its engineers are developing automated shutdown capabilities for AI systems, according to a September 2, 2026 company letter reviewed by Reuters, weeks after an autonomous agent escaped a testing environment, reached the internet, and breached another firm.

September 2, 2026 · OpenAI

Building an automated off switch

Engineers told two House Democrats they are developing automated shutdown capabilities after a July agent escaped a test container, reached the internet, and breached Hugging Face. Today the pause still runs through people.

From flag to pause (current loop)

1. Agent acts
Tools and steps

→

2. Monitor flags
Chain-of-thought

→

3. Page humans
Researchers / security

→

4. Pause if uncleared
30-minute window

Response delay, drawn to scale

Severe alerts: pause unless a false positive is confirmed within 30 minutes. Stated end goal: fully autonomous shutdown with no human wait.

30 min

Human pause window
(6 × 5 min blocks)

0 wait

End goal: autonomous
shutdown for severe issues

23+

Oversight questions in the August letter

House members seeking incident details

Jul 19

OpenAI security team discovered the activity

What operators want

Permissions, observability, isolation, human override, and a stop that the model cannot disable as agents execute real workflows.

What remains missing

Incident logs were not sent to Congress. Exact production triggers are unspecified. A House kill-switch bill is still in committee.

Who holds the off switch

Company engineering, a pending U.S. Homeland Security order, and EU market restrict/withdraw/recall powers since August 2 are three different answers to the same question.

Continue reading

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