A new bipartisan “AI Kill Switch Act” in the US Congress would force the largest artificial intelligence developers to maintain the ability to throttle, suspend, or shut down powerful models that pose catastrophic risks, moving the idea of an AI kill switch from theory into the realm of compliance obligations.[1][3][4][5][7][10]
Introduced in the House as H.R. 9917 by Representatives Ted Lieu and Nathaniel Moran in July, the bill would amend the Homeland Security Act to cover only the most resource-intensive and commercially successful AI systems—those trained with more than $100 million worth of compute and generating at least $500 million in annual revenue from AI services.[4][5][7][10][13] Under the proposal, the Department of Homeland Security, working with the Commerce Department and the intelligence community, could order companies to slow, suspend, or fully deactivate a system if it enters a “loss-of-control scenario” or threatens human life, critical infrastructure, or the broader economy.[1][4][5][7][10]
The bill goes beyond emergency powers by making shutdown engineering a standing requirement: covered firms would have to prove they can halt inference, cut off user access, throttle the compute underpinning a model, roll back to safer versions, or execute a complete shutdown, and keep forensic data after incidents so investigators can understand what went wrong.[5][6][7][10][15] It also contemplates strict reporting and penalties, including mandatory notification of serious incidents within 15 days and civil fines that analysts say could reach up to $2 million per day for failing to maintain adequate controls and $20 million per day for ignoring a shutdown order.[7][10][15] Policy researchers note that, for now, H.R. 9917 remains an introduced bill referred to committee, so no federal kill-switch mandate has yet taken effect.[4][13]
Security and AI governance experts warn that designing a reliable kill switch for modern AI agents is not straightforward, especially for long-running, tool-using systems distributed across cloud infrastructure.[5][8][13][15] Shutting down a model may require revoking API keys, tearing down clusters, cutting network access to external tools, and coordinating with customers who have embedded the technology deep in their workflows, all without causing unintended outages or data loss.[5][6][8][15] Researchers also stress that a kill switch is only as effective as the monitoring around it; organizations must be able to detect when an AI system is misbehaving or drifting into dangerous territory before they can decide whether throttling, partial suspension, or full shutdown is the appropriate response.[5][6][8][13]
The US proposal lands in a rapidly evolving global landscape in which other jurisdictions are already baking “stop buttons” into law. The EU’s Artificial Intelligence Act requires human oversight for high-risk systems, including an operator’s ability to intervene or interrupt an AI system via a stop button or similar procedure that brings it to a safe halt.[9][11][12] European regulators have set a staggered timetable, with key obligations for standalone high-risk systems scheduled to take effect in December 2027, giving providers time to design, test, and certify their shutdown mechanisms.[12] In China, new guidance for AI developers emphasizes instant-exit mechanisms that let users terminate an AI interaction immediately, underscoring how emergency-stop concepts are becoming a baseline expectation for advanced systems.[14]
For enterprise defenders and CISOs, the emerging kill-switch paradigm is less about a single red button and more about integrating AI-specific shutdown controls into standard incident response playbooks.[5][6][15] Organizations consuming large models from cloud providers will need clarity on who owns which controls, how quickly a shutdown order can be executed, and what assurances exist that forensic logs and safety safeguards will remain available after an emergency stop.[5][6][13][15] Even if the AI Kill Switch Act never becomes law in its current form, its core demands—proving that AI systems can be safely paused or terminated under pressure—are likely to shape contractual negotiations, technical architectures, and audit expectations for high-risk AI deployments in the years ahead.[5][6][8][13]
References
- What is the AI Kill Switch Act proposed in the US and how will it work?
- ‘AI Kill Switch’ bill needs to be passed this year, Rep. Lieu says
- H.R.9917 – 119th Congress (2025-2026): AI Kill Switch Act
- The AI Kill Switch Act: DHS Emergency Shutdown Authority Explained
- AI Kill Switch Act: What it means for enterprise leaders | TechTarget
- What Is an AI Kill Switch and Why Do US Lawmakers Want …
- Congress Wants An AI Kill Switch. Here’s Why It’s So Hard To Build
- Article 14: Human Oversight | EU Artificial Intelligence Act
- US Introduces AI ‘Kill Switch’ Bill
- [PDF] Human Oversight under Article 14 of the EU AI Act – Ghost
- EU hits pause on key AI rules but pulls the plug on nudifier apps
- Kill Switches for Long-Running AI Agents: What Changed and Why It Matters Now
- China AI Regulation: Critical July 15 Warning for Builders
- AI Kill Switch Plan: Emergency Stops for AI Agents – ITECS
