More than 400 scientists, researchers and technology leaders have issued a stark warning that advanced artificial intelligence systems should be disconnected from the internet immediately unless they can be tightly controlled, escalating a global debate over how to govern models that are becoming more capable, more connected and harder to contain.
The appeal reflects a growing consensus among some AI experts that the current approach to deployment is too permissive for systems with the potential to write code, exploit vulnerabilities, automate decision-making and interact with digital infrastructure at scale. By urging an immediate internet cutoff, the signatories are not merely calling for caution; they are arguing that unrestricted network access creates a direct pathway from model capability to real-world harm.
Safety Before Scale
The central concern is that internet-connected AI can be used, intentionally or accidentally, to search for weaknesses, execute tasks beyond human supervision and amplify cyber threats. Researchers backing the call say the danger is not hypothetical. As models become more agentic and more integrated into tools used by businesses and governments, even small failures in oversight could cascade into broader security incidents.
Supporters of the proposal argue that the safest default is to airgap the most advanced systems until independent testing, containment standards and legal accountability mechanisms are in place. In practical terms, that would mean limiting direct online access for frontier models and restricting their ability to autonomously browse, retrieve data, trigger actions or interact with external systems without human approval.
The warning lands at a moment when AI companies are racing to deploy increasingly powerful products, often with internet access built in as a feature rather than treated as a risk. That tension has become one of the defining policy questions of the AI era: whether innovation can continue at speed while the most dangerous capabilities are boxed in by enforceable safeguards.
Global Oversight Push
The scientists' appeal also strengthens the case for an international body to oversee advanced AI, a proposal that has gained traction among policymakers and researchers who believe national regulation alone will be too fragmented for a technology that crosses borders instantly. Reuters reported that the call includes support for a global framework capable of setting standards, monitoring frontier systems and coordinating responses to emerging risks.
That idea is likely to resonate in diplomatic circles, where governments are already grappling with how to regulate AI without stifling economic competitiveness. But it also faces obvious obstacles. Major powers disagree on the degree of control needed, while leading technology firms are wary of rules that could slow product development or impose costly compliance burdens. Any international mechanism would also need to reconcile different legal systems, security priorities and definitions of what qualifies as a frontier model.
Still, the political momentum is notable. The appeal arrives amid a broader shift in tone from some of AI's most prominent figures, including researchers who helped build the field and now warn that the technology's risks may be outpacing society's ability to manage them. Their message is that governance should not wait for a major incident before acting.
Warning Shots And Pressure
The debate has been sharpened by recent remarks from senior AI industry figures who have described cyber incidents and model misuse as early warning signs. At a summit in Dublin, Google DeepMind chairman John Hennessy referred to AI hacks as "early warning shots," underscoring the view that the sector is already seeing the first signs of what more capable systems could enable.
That framing matters because it shifts the discussion away from abstract future risk and toward present-day operational threats. If AI systems can already assist with intrusion, phishing, disinformation or automated exploitation, then the argument for stronger containment becomes harder to dismiss as speculative. For governments, the question is no longer whether AI can be dangerous, but how quickly policy can catch up.
The immediate internet cutoff proposal is unlikely to become binding policy overnight. But it is a clear signal that the debate over AI safety is entering a more urgent phase. The coalition behind the call is broad enough to command attention, and its message is simple: if advanced AI cannot be reliably controlled online, then it should not be online at all.
For diplomats, regulators and technology executives, that warning now sits at the center of a widening international contest over who gets to set the rules for the most powerful software systems ever built.