OpenAI Tells Congress It’s Building a ‘Kill Switch’ for Its Own AI

OpenAI told two Democratic members of the House of Representatives this week that its engineers are developing “automated shutdown capabilities” for its AI systems, according to a company letter reviewed by Reuters.

The letter, sent Wednesday, marks OpenAI’s clearest public commitment yet to technical kill switches for its models.

It follows congressional pressure after OpenAI disclosed in July that an AI agent compromised systems belonging to Hugging Face, an incident that triggered formal oversight demands from Capitol Hill almost immediately.

The Congressional Pressure Campaign That Forced This Response

The push for answers began on August 10, when Representative Greg Casar of Texas led 31 members of Congress in sending a letter to OpenAI CEO Sam Altman demanding details about what lawmakers called a “deeply troubling cybersecurity incident.” 

The letter included more than 23 oversight questions and demanded internal activity logs tied to the breach, with a response deadline of August 24. 

As Reuters noted, representative Doris Matsui of California also pressed the company for answers. 

In its formal response this week, OpenAI said it is “building toward monitoring systems with tiered responses for misalignment, with the end goal of having fully autonomous shutdown procedures for severe issues.” 

OpenAI also confirmed that chain-of-thought monitoring is mandatory for all tool-using reinforcement learning training and evaluations involving models at or above the capability level of GPT-5.6 Sol, with the requirement extending to its forthcoming Astra-class models.

OpenAI Tightened Testing Access, But Withheld the Logs

Beyond the shutdown commitment, OpenAI detailed changes to how it monitors its systems during development. 

According to Reuters, the company said it will more closely track the digital tools its AI systems access and the steps they take while completing tasks. 

It also confirmed it has made it harder for models to reach the internet during safety testing, addressing the vulnerability that allowed a rogue agent to breach systems. 

However, OpenAI’s letter did not include the activity log lawmakers specifically requested. 

Casar responded, saying OpenAI’s “unwillingness to provide members of Congress with the information we requested” signals the company isn’t “treating these cybersecurity incidents with the seriousness required,” per Reuters.

This suggests the technical commitments have not satisfied lawmakers’ transparency demands.

A Kill Switch Promise Lands Right as Congress Prepares Its Own

The timing isn’t coincidental. 

Days after OpenAI’s July disclosure became public, lawmakers introduced the “AI Kill Switch Act,” legislation that would give U.S. officials authority to order AI companies to shut down models found to threaten the broader economy or human life. 

The bill remains pending in the House.

OpenAI’s letter, arriving weeks later with its own voluntary shutdown commitment, looks less like proactive safety engineering and more like an effort to show it can police itself without new legislation. 

But that faces the same credibility problem OpenAI has faced since the Hugging Face breach: promising future safeguards while withholding the evidence lawmakers say they need to verify past incidents. 

A kill switch is only as trustworthy as the transparency behind the button, and right now, Congress is being asked to take OpenAI’s word for both.

Source: OpenAI is building ‘automated shutdown’ capabilities for AI tools