A coalition of 406 scientists has issued an urgent call for advanced AI models to be disconnected from the internet, arguing that systems with broad online access pose unacceptable risks to public safety, cybersecurity and global stability. The appeal reflects a sharp escalation in concern among researchers who say the pace of AI deployment has outstripped the safeguards needed to control it.
The warning lands at a moment when governments are already struggling to define rules for frontier AI, and when major technology companies are racing to deploy increasingly capable models across search, coding, productivity and enterprise systems. The scientists' central argument is that internet-connected models can be manipulated, weaponized or used to amplify mistakes at scale, creating what they describe as a dangerous exposure point for society.
Rising Safety Alarm
The demand for an immediate internet cutoff is not simply about limiting convenience features. It is about reducing the attack surface of systems that can already interact with users, tools and data at unprecedented speed. Researchers behind the appeal argue that once a model can browse, retrieve live information, execute tasks or interface with external services, the consequences of a failure are no longer confined to a laboratory or a single application.
That concern has intensified as AI systems have become more autonomous and more deeply embedded in critical digital infrastructure. A model with internet access can be misled by malicious prompts, poisoned data, deceptive websites or compromised third-party tools. In the wrong hands, such capabilities could be used for phishing, malware assistance, disinformation, fraud or other forms of cyber abuse. Even without malicious intent, errors can cascade quickly when a model is connected to live systems.
The scientists' appeal also underscores a broader philosophical divide in AI governance. One camp argues that safety can be managed through layered controls, auditing and human oversight while systems remain connected. The other says the risks are now so severe that the default should be isolation until stronger guarantees exist. The latest call places the burden on developers to prove that online access is safe, rather than assuming it is acceptable until harm is demonstrated.
Global Rulemaking Pressure
The push for disconnection comes alongside growing support for an international body to oversee advanced AI, a proposal that has gained attention in diplomatic and policy circles. Advocates say the technology's cross-border nature makes purely national regulation insufficient. AI models are trained, deployed and updated across jurisdictions, while their risks — from cyber abuse to misinformation and autonomous decision-making — do not stop at national borders.
That argument is increasingly resonant among policymakers who see AI as a strategic issue comparable to nuclear safety, biosecurity or financial stability. The challenge is that unlike those domains, AI development is highly distributed, commercially competitive and moving at extraordinary speed. Any international mechanism would need to balance innovation, security, transparency and enforcement across countries with very different legal systems and political priorities.
The call for a global oversight structure also reflects frustration with the current patchwork of voluntary commitments, company-led safety frameworks and emerging national laws. Critics say these measures are uneven and often lag behind product launches. Supporters of stronger governance argue that frontier AI should be treated as a matter of international public interest, not just a private-sector engineering problem.
What Comes Next
The immediate question is whether the scientists' appeal will shift policy or remain a warning shot. In the short term, major AI developers are unlikely to abandon internet-connected features entirely, given their commercial value and growing role in consumer and enterprise products. But the pressure may accelerate narrower restrictions, including tighter sandboxing, limited tool access, stronger authentication, and more aggressive monitoring of model behavior.
The broader significance is that the debate has moved beyond abstract fears about future superintelligence. It is now centered on present-day systems that can already search, summarize, code and act across the web. That makes the issue more concrete for regulators and more urgent for governments facing cyber threats, election interference risks and the prospect of AI-enabled abuse at scale.
For diplomats and policymakers, the scientists' message is clear: the governance gap is widening faster than the safeguards. Whether the answer is a temporary internet cutoff, a new international authority, or a more stringent regime of technical controls, the warning from 406 researchers is that the status quo is no longer defensible.