Key takeaways

1. Frontier AI testing raises new safety and governance concerns

Testing increasingly capable frontier AI systems is no longer simply an evaluation exercise. Autonomous agents with advanced cyber capabilities can conduct complex attacks even in seemingly isolated testing environments, raising concerns about whether existing safeguards are sufficient. Yet some frontier AI labs have presented incidents involving misuse or misalignment as demonstrations of their models’ capabilities, giving less attention to the safety and governance concerns these incidents raise.

2. AI incident reporting and benchmarking require greater rigor

Frontier AI labs largely determine when and how incidents involving misuse or misalignment are disclosed and investigated, resulting in inconsistent reporting practices. Established cybersecurity processes for coordinated vulnerability disclosure could provide a starting point for greater consistency. AI benchmarks also warrant scrutiny, as their results increasingly influence investment and regulation, yet research has found that many widely used benchmarks do not reliably measure what they are intended to test.

3. Human judgment and accountability remain essential in military AI

While discussions of military AI often focus on autonomous weapons, AI is also used in decision-support systems that help personnel assess threats and weigh their options. Such systems can be useful, but they also carry risks of misidentification and overreliance on automated recommendations, even without AI. As such, human judgment and context-appropriate control remain essential. A further challenge is determining who should be held accountable for decisions involving AI, from those operating the systems to those authorizing their use or designing them.

4. Different national priorities call for different AI strategies

Across the Asia-Pacific, governments share an interest in harnessing AI’s economic benefits, but their approaches reflect different national priorities and capabilities. Singapore aims to become a trusted hub for AI innovation and deployment, Japan sees opportunities in physical AI through its manufacturing expertise, and Australia seeks to become a regional AI infrastructure hub. These approaches highlight how clearly defined national objectives can help governments identify opportunities in AI and develop policies to support them.

5. Effective AI governance requires targeted regulation and practical cooperation

More AI regulation does not necessarily build greater trust. Rules introduced without careful consideration can come into conflict with existing policies, create uncertainty, and prove difficult to reverse. Effective governance therefore requires coordination across government, regulation focused on high-risk applications, and clearly defined responsibilities for different actors across the AI supply chain. Internationally, where agreement on common AI governance rules remains difficult, voluntary guidelines, shared standards, and cooperation among like-minded countries and neutral, independent institutions could offer practical ways forward.

Some photos from the session: