On 25 May in Canberra, the UK and Australia agreed a new frontier AI security arrangement centred on closer work between the UK AI Security Institute and the Australian AI Safety Institute. The Department for Science, Innovation and Technology said the Memorandum of Understanding is intended to improve how both countries identify and respond to risks from the most capable AI systems, with cyber security set out as an immediate priority. The agreement was signed by UK AI Minister Kanishka Narayan and Australia’s Assistant Minister for Science, Technology and the Digital Economy, Dr Andrew Charlton. The announcement places the work in a government-to-government setting, with the focus on public-interest research, system testing and technical coordination rather than a commercial partnership.
Under the arrangement, the two institutes are expected to exchange information on frontier AI capabilities, share research findings and work on common practice for testing and evaluation. The government statement also points to staff exchanges, which would give each side more regular contact on methods, evidence and risk assessment. That matters because evaluation standards for advanced AI systems are still being formed. Closer alignment between two allied institutes may make it easier to compare findings on model behaviour, misuse potential and the effectiveness of safety controls across jurisdictions.
The timing reflects a specific concern raised in the statement: advanced AI systems are improving quickly in their ability to support complex cyber-attacks. According to the UK AI Security Institute, recent research shows these capabilities are advancing rapidly, creating a two-sided policy problem. The same systems may help defenders identify weaknesses, process threat data and strengthen resilience. They may also reduce the skill or time needed for hostile actors to automate parts of an intrusion. That is why the agreement concentrates on security testing and evidence-sharing, rather than broad claims about AI innovation in the abstract.
The UK government said the partnership will also support work on international best practice for evaluating AI systems. In practical terms, that means developing more consistent ways to test whether a model behaves as intended, how it performs under misuse attempts and what safeguards are effective when systems are deployed in sensitive settings. For officials working on cyber policy, digital regulation and public-sector procurement, that is a material point. Testing methods often shape later decisions on assurance, access controls and whether high-capability systems can be used safely in government or critical national infrastructure contexts.
The Canberra agreement also extends work the UK institute is already carrying out with overseas research bodies. The statement says the institute shares practice through the International Network for Advanced AI Measurement, Evaluation and Science, alongside bilateral partnerships with other major economies. The Australian arrangement gives that work a more defined channel with a close ally that shares similar security interests. For the UK, the stated domestic purpose remains clear: research that informs policymaking aimed at protecting businesses, critical infrastructure and the public. Better access to Australian evidence and testing approaches could strengthen that work, particularly where model capability changes faster than one country can study on its own.
No new duties for businesses or AI developers were announced alongside the pact. The immediate effect is institutional: a formal route for the two institutes to share technical findings and develop testing practice together. For organisations outside government, the announcement does not set out new reporting, licensing or compliance steps. Even so, the arrangement may shape later policy choices. Common evaluation methods, shared research and routine staff contact can all feed into future decisions on cyber defence, procurement standards and how governments assess frontier AI systems before or after deployment.
Narayan described the pact as an extension of an existing UK-Australia security relationship and argued that no country can manage AI-related risks alone, particularly in cyber security. That position reflects a wider shift in AI governance, with governments placing more weight on technical evaluation, trusted international channels and faster exchange of evidence as model capability changes. The next stage is implementation. If the two institutes begin staff exchanges, publish joint research or align evaluation procedures, the Canberra agreement will move from a diplomatic statement to a working model for allied cooperation on frontier AI oversight.