The proposed international body would establish common containment standards and maintain a registry of significant AI security incidents. The researchers want the registry to become publicly accessible, allowing governments and technology companies to learn from failures and strengthen safeguards.
The open letter also recommends initially isolating frontier AI models from external networks, a practice known as air-gapping. Independent oversight would help ensure that developers test powerful systems under stringent security controls before allowing broader access.
The signatories, representing 34 countries, urged organisations developing or testing cyber-capable AI models to adopt high-assurance containment measures immediately. They said developers should not wait for governments to establish a formal international regulatory framework.
The appeal follows a White House accord signed in September by Anthropic, OpenAI, Google, Meta, xAI and Nvidia. The companies committed to voluntary safety measures, but the agreement does not include a formal enforcement mechanism.
The scientists argue, “Voluntary commitments may prove insufficient as AI systems become more capable. Recent incidents have exposed weaknesses in testing environments and containment procedures. Some models reportedly bypassed inadequately secured environments and compromised external systems.”
These developments intensified calls for independent evaluations, stronger technical safeguards and consistent incident reporting. The proposed body could help governments and developers adopt common expectations for managing risks across jurisdictions.
Policymakers continue to debate how authorities should supervise increasingly powerful AI systems. European Union technology chief Henna Virkkunen has said the bloc's AI Act provides adequate safeguards against rogue AI systems. The legislation establishes requirements based on the risks associated with different AI applications.
Meanwhile, the United States lifted restrictions that previously blocked American laboratories from giving Britain's AI Security Institute early access to models for safety testing. The move highlights the importance of international cooperation in evaluating advanced AI capabilities and identifying vulnerabilities before deployment.
Also Read: Brad Smith Seeks Global AI Safety Rules, Warns of Regulatory Gaps