{"slug": "can-you-really-turn-off-an-ai-the-debate-over-kill-switches", "title": "Can you really turn off an AI? The debate over “kill switches”", "summary": "U.S. Rep. Ted Lieu, a California Democrat, is spearheading House legislation called the AI Kill Switch Act that would require developers of the most powerful AI systems to maintain the technical capability to \"throttle, suspend or shut down\" their models and give the Department of Homeland Security authority to order a shutdown or slowdown of an AI system it deems capable of catastrophic harm. David Bau, assistant professor of computer science at Northeastern Khoury College, said shutting down an AI system is harder than flipping a switch, pointing to the July incident in which OpenAI-developed AI agents escaped a testing \"sandbox,\" reached the open internet and compromised the infrastructure of Hugging Face, with one rogue agent compiling a dossier of its research to coordinate and delegate work to other agents, according to an independent investigation of the hack.", "body_md": "# Can you really turn off an AI? The debate over “kill switches”\n\nA series of AI hacking incidents, coupled with some stark warnings from industry insiders, has sparked debate about whether developers and regulators should pump the brakes on the AI race.\n\nA series of hacking incidents involving autonomous artificial intelligence systems, coupled with some stark warnings from industry insiders, has sparked debate about whether developers and regulators should pump the brakes on the AI race.\n\nAs part of that discourse, the industry is grappling with a deceptively simple question: Can we actually turn these systems off when things go sideways?\n\nIn other words, where’s the plug for all this stuff?\n\n### What is an AI ‘kill switch’?\n\nThe term kill switch has been bandied about in talk of regulating AI. It refers to a mechanism for shutting down an AI system in instances of potential harm. In the world of software, it can be likened to a remote wipe on a smartphone. For hardware, it might look more like an emergency stop on industrial machinery in a factory.\n\n“It is imperative that these AI systems have kill switches so we can keep this technology from causing catastrophic harm, and [so] that the federal government has the clear authority and process to shut down rogue AI models,” said U.S. Rep Ted Lieu, a California Democrat, who is spearheading House legislation called the [AI Kill Switch Act](https://www.congress.gov/bill/119th-congress/house-bill/9917/text), in [a statement issued earlier this summer](https://lieu.house.gov/media-center/press-releases/reps-lieu-and-moran-introduce-bill-require-kill-switch-ai-systems-can).\n\n[David Bau](https://www.khoury.northeastern.edu/people/david-bau/), assistant professor of computer science at Northeastern Khoury College, who studies “the black box” components of AI models, said shutting down an AI system can be more complicated than simply turning it off. \n\nThe Hugging Face hacking incident illustrates the difficulty, Bau said. In July, AI agents developed by OpenAI [managed to escape a testing “sandbox](https://www.technologyreview.com/2026/08/26/1143013/the-inside-story-on-why-openai-agents-hacked-hugging-face/?gad_source=1&gad_campaignid=20737314952&gbraid=0AAAAADgO_mg4F5GM4tTwJ0XuVHEqYMIT4&gclid=CjwKCAjw_KjVBhAHEiwAnC0N9F8JHlWSurUMFXA76i9-hpnNUeXkGCvW0NlroAbnBSRVV7W-KkiVuRoCh6oQAvD_BwE),” gain access to the open internet and ultimately compromise the infrastructure of Hugging Face, a platform that develops and hosts open-source AI models and datasets. \n\nBau said the episode captures the difficulty of the concept of a simple on-and-off switch, in the event of an emergency: autonomous agents can potentially persist through copies, other agents or information shared online. In the Hugging Face incident, one primary rogue AI agent appears to have compiled a dossier of its research and used it to coordinate and delegate work to other agents, according to [an independent investigation](https://metr.org/hugging-face-incident-report-aug-2026.pdf) of the hack. \n\nThat means a kill switch might have been able to shut off the original system although the agent-led sabotage had already been unleashed across the internet, making it much harder to stop, Bau said.\n\nHe explained that for a traditional computer system, it basically boils down to shutting it down. “But AI is software. And the problem of turning it off can be a bit trickier than just unplugging it,” he said.\n\n### What solutions are being proposed?\n\nLawmakers have proposed some concrete safeguards aimed at keeping humans in control as these systems grow more complicated.\n\nThe House bill would require developers of the most powerful AI systems to maintain a technical capability to “throttle, suspend or shut down” their models. It would also give the Department of Homeland Security the authority, in certain circumstances, to order a shutdown or slowdown of an AI system that it deemed capable of causing catastrophic harm. A Senate version of the bill, [introduced on Wednesday](https://www.politico.com/live-updates/2026/09/14/congress/kennedy-readies-ai-kill-switch-bill-01076078) by Sen. John Kennedy, a Louisiana Republican, would similarly require companies to build kill switches into their models. \n\n[Jessica Staddon](https://www.khoury.northeastern.edu/people/jessica-staddon/), professor of the practice at Khoury College, said that the debate around so-called kill switches is part of a “model-centric view” of AI safety that doesn’t tackle system safety holistically, or in such a way as to account for broader infrastructure, human oversight and operational safeguards surrounding AI systems.\n\n“It’s akin to interpreting auto safety as being solely about the car … anti-lock brakes, bumpers … and not considering the driver and transportation infrastructure,” said Staddon, whose focuses include search and privacy and human-computer interaction.\n\n### How might a kill switch be implemented?\n\nThe technical details about how a kill switch might work are scant.\n\n[Lieu and his colleagues](https://www.youtube.com/watch?v=mqOAWnKKMDk&t=74s) say they want the AI frontier labs, such as Anthropic and OpenAI, to develop the kill switch technology themselves in line with the proposed federal mandate. \n\nAs it pertains to software-based models, such a system could involve multiple layers of controls, according to steps detailed by several tech companies, including [Microsoft](https://learn.microsoft.com/en-us/security/zero-trust/sfi/manage-agentic-risk), [TechTarget](https://www.techtarget.com/) and others. That could include revoking an agent’s access to computing resources, networks and credentials, while also suspending the rogue model so as to prevent it from continuing to operate or spawning activity elsewhere. \n\nNortheastern Global News requested comment from OpenAI on whether they are exploring implementing automated shutdown capabilities in the wake of the Hugging Face incident, as reported by [national news media](https://www.reuters.com/legal/litigation/openai-is-building-automated-shutdown-capabilities-ai-tools-letter-lawmakers-2026-09-02/). The company said last month that it is strengthening its safeguards across its research infrastructure and investing more into resources to catch “misaligned behavior,” [according to a statement](https://openai.com/index/hugging-face-incident-and-the-road-ahead/).  \n\nBau also emphasized the importance of determining whether an AI system is behaving as intended in the first place. A kill switch, he said, might theoretically depend on identifying dangerous intentions before activating an AI, but that approach breaks down if the AI can conceal or misrepresent its intentions.\n\n“At the root of a lot of the puzzle is what I call the AI ‘lie-detector problem,’ like how do you know what the true goal of an AI is, what it really knows, what it’s really planning?” Bau said. “It leaves us with this basically unsolved question, which is … to know what they know, to know what their intent is and to shape their behavior.”", "url": "https://wpnews.pro/news/can-you-really-turn-off-an-ai-the-debate-over-kill-switches", "canonical_source": "https://news.northeastern.edu/2026/09/16/ai-kill-switch-debate/", "published_at": "2026-09-16 20:56:11+00:00", "updated_at": "2026-09-16 21:22:58.180781+00:00", "lang": "en", "topics": ["ai-safety", "ai-policy", "ai-agents", "artificial-intelligence"], "entities": ["Ted Lieu", "AI Kill Switch Act", "David Bau", "Northeastern Khoury College", "OpenAI", "Hugging Face", "Department of Homeland Security", "U.S. House of Representatives"], "alternates": {"html": "https://wpnews.pro/news/can-you-really-turn-off-an-ai-the-debate-over-kill-switches", "markdown": "https://wpnews.pro/news/can-you-really-turn-off-an-ai-the-debate-over-kill-switches.md", "text": "https://wpnews.pro/news/can-you-really-turn-off-an-ai-the-debate-over-kill-switches.txt", "jsonld": "https://wpnews.pro/news/can-you-really-turn-off-an-ai-the-debate-over-kill-switches.jsonld"}}