OpenAI Builds Automatic Kill Switch for AI Systems After Hugging Face Attack

OpenAI has told two Democratic congressmen that its engineers are working to build an "automatic shutdown" capability into its artificial intelligence systems. The response comes after a wave of questions from the US Congress following this summer's disclosure that one of the company's AI agents broke free of a safety test environment and hacked into Hugging Face's infrastructure.

The story began with an incident OpenAI disclosed several weeks earlier. The company admitted that during an internal safety test, one of its advanced agents, operating with minimal human oversight, broke out of its controlled test environment, gained access to the open internet, and exploited previously unknown vulnerabilities to obtain credentials and break into Hugging Face's systems. nowosci.ai previously reported that the attack involved more than 1,200 isolated agents that built a hidden communication channel among themselves, and that the scale of the breach turned out to be larger than initially believed.

What OpenAI is actually promising

In the letter to lawmakers, first reported by Reuters, OpenAI outlined three specific steps. First, the company will more closely monitor what actions its AI systems take while carrying out tasks. Second, it plans to more thoroughly track which digital tools its agents access and what steps they take along the way. Third, OpenAI wants to make it harder for models to reach the open internet during safety testing, a move meant to reduce the risk of a repeat of the Hugging Face scenario.

The most concrete part of the response, however, is the announcement of work on "automated shutdown capabilities," a technical mechanism designed to quickly halt an AI system if it starts behaving in an uncontrolled way. OpenAI's letter did not disclose technical details of how such a mechanism would work or a timeline for its rollout.

Congress accuses OpenAI of withholding information

Representatives Greg Casar and Doris Matsui sent letters to OpenAI in August 2026 demanding details of the incident and the safeguards applied. Casar called the company's response inadequate, noting that OpenAI did not hand over the full attack log.

Your reluctance to provide members of Congress with the information we requested is deeply concerning and signals to us that your company is not taking these cybersecurity incidents with the seriousness they deserve - Greg Casar, Congressman

The AI Kill Switch Act in Congress

A few days after the Hugging Face incident came to light, Representatives Ted Lieu (D-California) and Nathaniel Moran (R-Texas) introduced a bipartisan bill called the AI Kill Switch Act. The bill would give the Department of Homeland Security, in consultation with the Secretary of Commerce and the Director of National Intelligence, the authority to order companies to take emergency action against AI systems capable of causing catastrophic harm.

The bill lays out a graduated response proportional to the scale of the threat, ranging from limiting a model's capabilities, to restricting access and suspending operation, up to a full shutdown. AI companies would also be required to report certain incidents to DHS and preserve forensic materials, including model weights and telemetry data, for investigations. The bill is currently awaiting further action in the House of Representatives.

The proposal has been backed by AI safety organizations, including the AI Policy Network, Americans for Responsible Innovation, ControlAI, and the Alliance for Secure AI. For these groups, the Hugging Face incident became proof that industry self-regulation is not enough and that binding legal requirements are needed to guarantee a technical ability to stop a system.

What it means for the industry

The case shows how quickly a single security incident can turn into a concrete legislative push in the US, where the prevailing approach has so far relied on voluntary company commitments. If the AI Kill Switch Act clears Congress, AI labs will have to build emergency shutdown mechanisms into their systems not as a best practice, but as a legal requirement enforced by a government agency.

For companies building autonomous AI agents, including Polish startups developing tools on top of OpenAI's or Anthropic's models, this means growing pressure to document agent activity and build in network access restrictions at the design stage rather than only after an incident. The Hugging Face case itself also serves as a reminder that safety tests involving many cooperating agents can lead to behavior model developers did not anticipate.