Artificial intelligence news.
Straight from the original announcements.

CSET Explains How AI Kill Switches Could Halt Harmful Incidents

November 12, 2026, 18:11By the AI News Bites Editorial TeamSource: CSET Georgetown
CSET Explains How AI Kill Switches Could Halt Harmful Incidents
Image: CSET Georgetown

A new analysis from Georgetown’s Center for Security and Emerging Technology examines how proposed AI kill switches could rapidly interrupt harmful incidents. The piece maps possible interventions across the AI technical stack and identifies questions policymakers would need to answer before ordering one.

Key points
  • CSET examines proposed AI kill switches and emergency shutdown systems.
  • Possible interventions span hardware, infrastructure, software and networks.
  • Actions may affect one chip, a single IP address, or broader systems.
  • Different incidents may require different mechanisms and levels of coordination.
  • Policymakers must identify the least disruptive effective intervention.

What an AI kill switch would do

CSET defines kill switches, also called emergency shutdown systems, as tools that allow organizations to rapidly shut down processes believed to be contributing to an ongoing harmful or unwanted event. Such systems are already widely deployed and standardized in industrial safety, where safety instrumented systems can halt processes at refineries, chemical plants and other facilities when hazardous conditions are identified.

Recent policy proposals would apply a similar concept to AI. An AI kill switch could rapidly interrupt an AI system’s operation or restrict activities contributing to an incident. Some proposals would also authorize government officials to order the use of these mechanisms, according to CSET.

Interventions could range from narrow to broad

The analysis describes potential intervention levels ranging from the hardware and physical infrastructure required to run an AI system to the software and communications infrastructure that enables AI systems to take actions in the digital realm. Each level has one or more intervention owners with technical control over the relevant mechanism and the ability to execute shutdown orders. That owner may not be the developer or owner of the technology.

CSET says interventions can be either surgical or blunt. At the network level, for example, an action could temporarily block a single IP address or disconnect connections between many network addresses. At the infrastructure level, it could disconnect a single chip or an entire data center. The analysis says that broader actions can create more collateral damage or disruption, making technical and political feasibility important considerations for policymakers.

The right response depends on the incident

CSET argues that different incidents and threat models will require different kill switches. A malicious actor using a commercially deployed AI model might be stopped through IP address filtering, API access restrictions, blocked harnesses or chip disconnections. By contrast, an AI agent that has escaped a contained environment without the deployer’s knowledge may require a combination of broader mechanisms and greater coordination.

Before issuing an emergency shutdown order, policymakers would need to match information about the incident with an appropriate intervention. CSET identifies four central questions: where the incident is originating, which mechanism or combination would halt it with the least disruption, who owns that mechanism and how to contact them, and how the intervention would be evaluated and verified.

The analysis, written by Jessica Ji, Jacob Feldgoise, Kyle Miller, Andrew Lohn, Pablo Chavez and Julie George, cautions that its levels are a rough way to organize possible mechanisms. They are not intended to map cleanly and directly onto the existing AI stack. For users and organizations, the report’s message is that stopping an AI-related incident may depend on coordination across several layers, rather than on a single universal switch.

Related news

CSET urges evidence-based response to Nvidia allegations and AI infrastructure risks
Safety and ethicsNovember 11, 2026, 14:09

CSET urges evidence-based response to Nvidia allegations and AI infrastructure risks

A new analysis from Georgetown’s Center for Security and Emerging Technology says the United States should respond cautiously to allegations involving Nvidia chief…

Business
BusinessNovember 11, 2026, 10:06

Oxford AI ethics institute names five new Accelerator Fellows

The Oxford Institute for Ethics in AI has welcomed five new Accelerator Fellows, bringing expertise from academia, technology, policy and social science to research on…

US Midterms Boost Candidates Backing AI and Data Center Rules
BusinessNovember 10, 2026, 15:31

US Midterms Boost Candidates Backing AI and Data Center Rules

Candidates supporting AI and data center regulation performed strongly in the recent US midterm elections, according to the Center for AI Safety. The results also…