Singapore, Microsoft to explore ways to test the safety of frontier AI models

Singapore’s Infocomm Media Development Authority (IMDA) and Microsoft have agreed to explore ways to test frontier AI models for safety and security under a Memorandum of Understanding signed on June 12. The collaboration follows global concerns over Anthropic’s Claude Mythos Preview, which autonomously discovered thousands of high-severity vulnerabilities in major operating systems and web browsers, some hidden for decades. The model was restricted under Anthropic’s Project Glasswing program and prompted urgent warnings from U.S. Treasury Secretary Scott Bessent and Federal Reserve Chair Jerome Powell to bank executives about AI risks. The partnership will focus on assessing AI models, including their safe operation across different languages, and sharing best practices for governance. IMDA highlighted the need for collaboration as AI development outpaces individual efforts, emphasizing a ‘trusted ecosystem’ to balance innovation with security. Deputy CEO Kiren Kumar stated the collaboration demonstrates how government and industry can work together to scale responsible AI governance. Microsoft’s Chief Responsible AI Officer, Natasha Crampton, noted Singapore’s role in shaping global AI safety standards. Anthropic’s Claude Mythos Preview identified over 10,000 flaws, many of which are now being remediated, according to IMDA. The authority suggested early access to AI models by trusted partners could help preempt AI-enabled exploitation. While Microsoft is not a primary developer of frontier AI models like Anthropic or OpenAI, it provides access to multiple models through Azure AI Foundry, including those from OpenAI and Anthropic. Singapore’s Cyber Security Agency previously urged firms to patch critical vulnerabilities and implement multi-factor authentication after Anthropic’s testing. The collaboration aligns with Singapore’s broader approach to AI governance and follows similar agreements Microsoft has signed with the UK’s AI Security Institute and the U.S. Centre for AI Standards and Innovation. IMDA’s goal is to ensure AI development remains secure, reliable, and beneficial to the public.