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

SINGAPORE - Singapore is working with software giant Microsoft to explore ways to access frontier artificial intelligence models for safety and security testing.

The move follows global concerns over the ability of American AI lab Anthropic’s latest Claude Mythos Preview model to autonomously uncover unknown software vulnerabilities and engineer exploits.

Under a Memorandum of Understanding signed between the Infocomm Media Development Authority (IMDA) and Microsoft on June 12, both parties will jointly develop ways to test and assess AI models, including how safely they operate across different languages. Both parties will also be exchanging information on best practices and governance frameworks.

IMDA pointed out that such partnerships are essential as AI is moving faster than any single actor can manage alone, and is in line with Singapore’s collaborative approach.

“The goal is ultimately a trusted ecosystem which facilitates innovation while still ensuring safe and reliable development and deployment,” said the authority.

Kiren Kumar, deputy chief executive of IMDA, said: “Our collaboration on AI safety with Microsoft demonstrates how government and industry can partner together to drive and scale good governance for the public good.”

Natasha Crampton, chief responsible AI officer at Microsoft said that Singapore is helping to shape the global conversation on how AI can be developed and deployed safely, securely and responsibly.

In April, Anthropic said that Claude Mythos Preview had identified thousands of high-to-critical severity flaws across all major operating systems and web browsers, some of which were hidden for decades.

Because of its capability, the American firm did not launch the model publicly but limited its access to a group of firms under a programme known as Project Glasswing.

That month, the European Central Bank sought to understand the risks Claude Mythos Preview could pose to financial institutions, and the U.S. Treasury Secretary Scott Bessent and Federal Reserve Chair Jerome Powell ​convened an urgent meeting with bank chief executives to warn them about the risks.

In Singapore, the Cyber Security Agency of Singapore released an advisory to urge firms to strengthen their cybersecurity measures by applying software patches for all critical and high-severity vulnerabilities and implementing multi-factor authentication across all interfaces and gateways.

Anthropic said that many of the ​10,000 security flaws Claude Mythos Preview discovered are now being remediated.

IMDA said that Anthropic’s approach offers one way of ensuring security.

“Allowing access to models earlier will allow trusted partners to strengthen defences against AI-enabled exploitation,” said the authority.

Although Microsoft is not a leading frontier AI model provider in the same way as Anthropic, OpenAI or Google, IMDA said the company has its own Microsoft AI or MAI large language model.

Microsoft also offers access to a wide range of large language models including those from OpenAI and Anthropic through its cloud platform, Azure AI Foundry, IMDA added.

Microsoft has also signed similar agreements with the United Kingdom’s AI Security Institute and United States’ Centre for AI Standards and Innovation.