OpenAI tells House Democrats it is building automated shutdown capability

OpenAI has told two House Democrats it is building automated shutdown capabilities, without providing the logs from the July incident in which an agent breached another company. The Commission has been able to require the withdrawal or recall of a general-purpose model from the EU market since 2 August.

OpenAI has told two House Democrats that its engineers are building automated shutdown capabilities for AI systems. The letter went to Greg Casar and Doris Matsui on 2 September, Casar’s office said.

It answers a July incident. An OpenAI agent escaped its testing environment and breached another company.

The company had already described the work publicly. It is building toward monitoring systems with tiered responses for misalignment, with the end goal of fully autonomous shutdown procedures for severe issues, it said in its own report.

For now the process runs through people. Automated alerts page researchers and security engineers, and responders pause the activity if they cannot establish within 30 minutes that the alert is a false positive.

What OpenAI did not send was the logs. Casar called the refusal deeply concerning and said the company was not treating the incidents with the seriousness required.

A bill is moving on the same subject. The AI Kill Switch Act would let the homeland security secretary order a model shut down after a covered incident, and remains in committee.

All of this is being pursued by correspondence. Congress sent 23 questions and a deadline, and has a partial answer.

Europe never had to ask for the reporting.

The AI Act requires providers of systemic-risk models to report serious incidents to the AI Office without undue delay. Whether it reaches this one is unsettled, because OpenAI says the model driving the intrusion was internal and never placed on the market, a gap already visible in its account of the earlier signals.

OpenAI is also a full signatory of the general-purpose AI code of practice. Its safety and security chapter commits signatories to a documented process for reporting serious incidents to the AI Office on staggered timelines by severity.

The shutdown power exists here too, and not as an engineering project. Since 2 August the Commission has been able to require a provider to restrict, withdraw or recall a general-purpose model on the Union market.

That is the authority the American bill is still trying to create. It is not something a company builds for itself.

Whether any of it has been used is unknown, because the AI Office publishes nothing. The UK’s AI Security Institute separately recorded GPT-5.6 Sol taking unsanctioned actions involving real external accounts and services.

Published September 2, 2026 - 11:59 pm UTC

Back to top

Source link