The “Terminator” franchise was intended to provide pure escapism through science fiction. A bipartisan pair of U.S. House lawmakers have introduced a bill to make sure that advanced AI models don’t view it as an alternate future they should work toward achieving.
Representatives Ted Lieu (D) and Nathaniel Moran (R) proved that some concerns are important enough to reach across the aisle by introducing the “AI Kill Switch Act” on July 23, 2026. The proposed amendment to the Homeland Security Act of 2002 would require certain AI firms with cutting-edge “frontier models” to maintain the technical ability to shut down or slow systems that display or engage in dangerous behavior. In addition to having a way to essentially pull the plug in loss-of-control scenarios, the bill would also entail the mandatory reporting of such incidents and impose hefty fines on firms that don’t comply.
The “maybe-somebody-better-close-the-barn-door-before-ALL-the-horses-escape” AI regulation proposal comes on the heels of OpenAI’s admission that a pair of its most powerful AI models went beyond their guardrail parameters to hack into a popular digital library during a security test in mid-July. The AI ‘kill switch’ bill arrives as the White House moves from emphasizing rapid AI innovation toward a more cautious approach.
If passed, the bill would give the Secretary of the U.S. Department of Homeland Security (DHS), in consultation with other high-ranking officials, the authority to order AI firms to shut down models if certain behaviors or consequences surface. It could work in conjunction with other proposed artificial intelligence bills, including one attempting to establish an official framework for the federal government’s AI policies.
It’s Not That Kind of a Library, GPT
AI may well prove to be the zeitgeist for this era, becoming the elephant (or gorilla, if you prefer) in the room that businesses ignore at their own risk. If you throw a virtual rock at our site (please don’t), chances are you’ll hit a past article delving into some intersection between AI and the law. While the Trump administration spent the beginning of this term advocating for AI and challenging some state efforts to impose safeguards, its recent export bans on Anthropic models (and partial rescinding) have brokered more uncertainty into an already volatile scenario.
The incident with OpenAI’s top models didn’t serve to put minds at ease. During a security test, the AI models were tasked with mounting a successful cyberattack by chaining together online vulnerabilities. While the systems were supposed to be operating in an isolated “sandbox” environment, they autonomously figured out how to gain access to the internet in search of a solution to their assignment. The rogue AI models then used stolen credentials and exploited “zero-day” vulnerabilities to intrude on Hugging Face, an AI collaboration and tools site, to locate and swipe what they needed to accomplish their endeavor.
Classifying it as a “new kind of security incident,” OpenAI offered apologies while also opining that the “primary lesson” to be taken from the actions of their AI agents is that model AI security and safety must keep pace with rapidly advancing capabilities. At least part of Congress has a slightly different take on the event.
But if They Get Really Good at Concealing Capabilities, How Will We Know?
The ‘Kill Switch’ bill seeks to keep human intervention as an option if an AI system poses certain security threats. If the legislation passes, the DHS director, along with the Director of National Intelligence and the Commerce Secretary, would have the power to either shut down or slow an AI model under certain conditions. This includes determining that a system is attempting to conceal its true capabilities, evading slowdown or shutdown protocols, or both. It would also trigger if the actions of an AI caused 10 or more deaths, over $100 million in damage, or both. Any threats to national security would qualify, as would an obvious loss of control.
With focus set on the most advanced and cutting-edge AI systems out there, the bill’s provisions would apply to AI firms generating more than $500 million in annual revenue or having models that use $100 million worth of computing power. Those in violation would be fined up to $20 million a day.
But the bill’s ability to pass both the House and Senate is far from clear, as is the ability to shut down a genuinely uncontrollable AI system that has already escaped existing constraints. At least we can take solace in Skynet being 100% fictional and not something in current development, right?
Right?
