It was a privilege to deliver the opening keynote at the Australian AI Safety Forum, bringing together researchers, industry and government to discuss the findings of the International AI Safety Report and what they mean for Australia’s AI safety future.

My keynote argued that AI safety is entering a new phase. First, practical AI capability is increasingly determined at the system level. Harnesses comprising tools, verifiers, memory, autonomous feedback loops and operating environments increasingly determine what an AI system can actually do. These system-level capabilities are becoming both a key source of competitive advantage and intellectual property, while also introducing new safety risks.

Second, the frontier is shifting from pre-deployment benchmarking towards runtime verification, guardrails and continuous control.

Third, understanding AI risks increasingly depends on evidence from real-world deployment rather than laboratory benchmarks alone, making trusted mechanisms for analysing operational data essential to improving evaluations, safeguards and policy.

I was particularly encouraged to hear Minister Andrew Charlton highlight AISI’s collaboration with CSIRO on AI alignment research in his opening address. As he noted, alignment is fundamentally about ensuring AI systems do what we intend, not simply what we specify. Building the science and engineering needed to evaluate, verify and continuously control increasingly capable AI systems will be central to achieving that goal.

Thanks to the organisers, speakers and participants for an excellent forum, and to everyone contributing to Australia’s growing AI safety research and engineering community.

See slides: https://www.linkedin.com/posts/limingzhu_future-of-ai-safety-science-activity-7480361670263853056-zo94