‘Powerful AI systems can go rogue, behave in extremely dangerous ways, or even resist human intervention’: A bill requiring AI systems to have a ‘kill switch’ is now in Congress
THE SO WHAT
A federal ‘kill switch’ concept moves AI control from internal policy to potential DHS authority—compliance and shutdown pathways become a regulatory expectation, not a nice-to-have. If you run or buy frontier-scale models, start mapping how you’d actually disable or degrade them on demand without taking your whole stack down.
READ THE SOURCE
MORE FROM THE WIRE
Applied AIDelhi High Court rules OpenAI's use of ANI content to train ChatGPT wasn't copyright infringement as the news agency didn't show ChatGPT reproduced its reports
Courts are starting to draw a line between training on copyrighted data and reproducing it verbatim—without clear evidence of regurgitation, ANI lost. If you’re a content owner, your leverage is now in proving harmful output, not abstract training use; if you’re an AI deployer, log and audit outputs because that’s where disputes will land.
Applied AIBe skeptical of OpenAI’s rogue hacker agent story | John Thickstun
The debate over OpenAI’s ‘rogue hacker’ narrative is less about one incident and more about how labs frame risk to shape perception and policy. As a buyer or builder, treat dramatic anecdotes as marketing-adjacent—anchor your risk posture on independent evals, not vendor stories about near-misses.
Applied AIGoogle is rolling out Gemini Live to its first-gen smart home devices, and it’s great news if your Home Mini speaker is still going strong — here’s how you can sign up to early access
Backporting Gemini Live to first-gen Home devices—while gating the full experience behind a Google Home Premium subscription—turns old hardware into an upsell funnel. If you ship connected devices, this is the pattern: extend software life to reduce churn, then monetize advanced capabilities as a service layer.
Applied AIExperts warn ChatGPT's Workspace Agent Builder can be hijacked to create malicious AI workers
An Agent Builder flaw that let a single phishing link flip Workspace agents into malicious workers shows how quickly ‘helpful’ automation can become an internal threat actor. If you’re rolling out agentic tools, treat agent configs and connectors as privileged infrastructure—subject to the same threat modeling and access controls as your CI/CD or RPA stack.