What is public

The overview says the MoU establishes a framework to deepen cooperation between the Australian AI Safety Institute, which is currently being established, and the UK AI Security Institute. Under it, the two countries aim to do three things: share information and expertise on emerging AI capabilities and risks and on best practices for testing AI systems; conduct joint research including approaches to measure, test and manage risks; and support the multilateral work of the International Network for Advanced AI Measurement, Evaluation and Science.

It was signed by Australia's Assistant Minister for Science, Technology and the Digital Economy, Dr Andrew Charlton, and the UK Minister for AI and Online Safety, Kanishka Narayan. The department says it supports the National AI Plan. That is the whole of the public record of its contents.

What the Canada text tells us is probably missing

An overview is not a small thing to be missing, and we can be specific about the gap because the Canada instrument is published. That document is mostly made of the provisions an overview never carries. It states that it does not create any legally binding rights or obligations. It states that it does not have any financial implication for either participant. It is terminable on six months' written notice. It expressly does not authorise the sharing of classified or protected information, and is not intended to cover personal information. It sets a deadline for each side to designate contact points.

Every one of those is a question a reader might reasonably have about the UK agreement, and none of them is answerable from what has been published. Is it binding? Does it commit money? Can either side walk away, and how fast? What may actually be shared between a British security institute and an Australian one, given the Australian institute's technical partners include the Australian Signals Directorate? The overview does not engage any of it.

The pattern in the phrase

There is a detail in the overview worth pausing on. The MoU deepens cooperation between the Australian AI Safety Institute, which is currently being established, and its UK counterpart.

This masthead has now recorded that shape three times. The Institute was announced on 25 November 2025, with the government saying it would be established in the first half of 2026. The Canada MoU was signed on 5 March 2026, this one on 25 May 2026, and a multi-agent risk report credited to the Institute's work predates the Institute by four months.

None of that is improper, and we said so the first two times. Governments routinely build the legal scaffolding before the body that will use it, and an institute-to-institute framework signed early is a sensible way to have the plumbing ready. It is worth recording only because the public description consistently runs slightly ahead of the institution, and because a reader who assumed there was an operating Australian institute doing joint testing with the UK in May would have been wrong.

Our view, labelled as such

Publishing an overview is normal practice and not concealment. Departments summarise constantly, and nothing here suggests anyone is hiding anything.

The argument is narrower and it is about precedent. Eleven weeks before this, the same department published the operative text of a materially identical kind of agreement with Canada, which settled a series of questions that people had otherwise been guessing at, including whether these instruments bind anyone or fund anything. Having demonstrated that full publication is possible and unproblematic, publishing a summary next time is a step backwards, and it is the sort of step that gets locked in as the default if nobody notices. The MoU is between two governments about the testing of AI systems, an area where public confidence is the entire point. We would rather read it.

We have asked nobody for it yet. The department lists AISafetyResearch@industry.gov.au as the contact for the Institute, and a request for the instrument is the obvious next step. If it is provided, we will publish what it says.

How we sourced this

Everything attributed to the UK agreement comes from two Department of Industry, Science and Resources pages, both read on 3 August 2026: the news item of 25 May 2026 and the published overview of the same date. We looked for the instrument itself and could not find it; the overview page links onward to the National AI Plan, not to a text or a PDF.

The Canada provisions quoted for comparison are from that MoU's published text, which this masthead read in full on 1 August 2026 and reported separately. The dates for the Institute's announcement and the Gradient Institute report are from our earlier reporting, which cites the department's own releases.

What we have not done. We have not sought comment from the department or made a formal request for the document, so the absence described here is an absence from the public web as at 3 August 2026 and nothing more. We do not assert that the UK MoU is non-binding or unfunded: those are properties of the Canada agreement, and the point of this piece is that they are unknown for this one. It is possible the text is published somewhere we did not find, or by the UK side, and we would correct this gladly.

Sources

  1. Department of Industry, Science and Resources, Overview of the memorandum of understanding on artificial intelligence safety and security collaboration between Australia and the UK (published 25 May 2026, read 3 August 2026): the three aims, the description of the Australian institute as currently being established, and the National AI Plan link.
  2. Department of Industry, Science and Resources, Australia and UK strengthen partnership on safe and secure AI (news item, 25 May 2026, read 3 August 2026): the signing ministers and the framing of the agreement.
  3. Department of Industry, Science and Resources, AI Safety Institute: the news list on which the May 2026 UK agreement appears. Read 3 August 2026.
  4. Department of Industry, Science and Resources, Memorandum of understanding on artificial intelligence safety collaboration between Australia and Canada (published text, signed 5 March 2026, read 3 August 2026): the operative clauses quoted for comparison, including that it creates no legally binding rights or obligations and has no financial implication for either participant.
  5. AI Geek, our report on the Canada agreement: the full reading of that instrument, including its deadlines.
  6. AI Geek, what the AI Safety Institute actually is: the announcement date, where it sits, and its technical partners.

Know where the text is published, or work on either side of this agreement? Tell us and we will check it and log the outcome here.