AI News · The messy stuff ·
US and California initiatives put AI emergency controls on the policy agenda
TechTarget reports that frontier AI developers face competing safety requirements as policymakers address agentic systems. The Kill Switch Act, HR 9917, is pending in the US House of Representatives. California Governor Gavin Newsom's Executive Order N-9-26 directs the state to explore requiring emergency kill switches and greater independent oversight for AI developers.
Key points
- The Kill Switch Act, HR 9917, is pending in the U.S. House of Representatives.
- California’s executive order directs exploration of emergency kill switches and greater independent oversight for AI developers.
- Experts warned that conflicting technical requirements could increase compliance costs across markets.
- Security leaders advocated continuous control testing as models, tools and permissions change.
- Teams should distinguish proposed requirements from binding rules when evaluating shutdown controls and oversight.
What happened: Emergency shutdown controls for advanced AI are gaining attention in Washington and California, but the latest initiatives are not equivalent to settled requirements. TechTarget reported that the Kill Switch Act, HR 9917, is pending in the U.S. House of Representatives. California Governor Gavin Newsom’s Executive Order N-9-26 directs the state to explore requiring AI developers to build emergency kill switches and undergo greater independent oversight. For business teams building or adopting AI agents, the distinction matters: a pending bill and an instruction to explore requirements should not be treated as finalized technical obligations.
The details: The initiatives address agentic AI, where automated agents take on increasingly autonomous workflows. Emergency shutoffs are intended to provide a way to stop those operations, while independent oversight would test whether developers’ safety controls work. TechTarget reported that California is exploring independent verification of frontier AI companies’ safety frameworks and regular testing of emergency shutoffs. The wider policy landscape includes the AI Emergency Button Act, recently introduced in the Senate, and an Oregon executive order directing criteria for third-party reviews and contract kill switches. Illinois has established an AI Cabinet to evaluate shutdown mechanisms, while Virginia has launched a rapid-response AI Task Force.
Background: The challenge extends beyond U.S. state and federal boundaries. According to TechTarget, Article 14 of the European Union AI Act establishes mandatory human oversight for high-risk AI, with requirements phasing in through 2027. China’s AI Safety and Governance Framework 2.0 mandates circuit breakers and one-click control for autonomous operations. Sonatype CEO Bhagwat Swaroop told the outlet that differing state laws can be manageable when they seek the same outcomes, but conflicting technical measures drive costs. He compared the situation with data breach notification laws, where companies built toward the strictest standard rather than creating separate compliance programs for every state.
Who it affects: For developers and enterprise security teams, the issue is not simply whether a shutdown control exists, but whether it remains effective as systems change. Sahil Agarwal, field AI security fellow at Anaconda and founder of Enkrypt AI, told TechTarget that annual audits cannot keep pace when models, tools and permissions change weekly. Vectra AI CEO and president Hitesh Sheth argued that enforcement must become more continuous and evidence-based, including evidence of risk evaluation, control testing, incident monitoring and what happened when something went wrong.
What to watch: Teams building agents should evaluate shutdown controls and oversight processes now while tracking which measures become binding. Experts interviewed by TechTarget favored demonstrable safety outcomes over rigid technical specifications. Agarwal advocated requirements such as incident reporting and independent verification, with technical details left to standards that can evolve faster than legislation. The practical question is whether emerging rules converge on compatible outcomes or impose competing designs across markets.
Our take
Teams building agents should evaluate shutdown controls and oversight processes now, while distinguishing proposed requirements from rules already in force.