CONROY, Dr Kate, General Manager, AI Safety Institute, Department of Industry, Science and Resources GRAHAM, Mrs Michele, Acting Head of Division, Technology and Digital Policy Division, Department of Industry, Science and Resources HICKIN, Mr Lee, Executive Director, National AI Centre, Department of Industry, Science and Resources WILSON, Mrs Helen, Deputy Secretary, Science and Technology Group, Department of Industry, Science and Resources [09:49] CHAIR: Welcome. I understand that information on parliamentary privilege and the protection of witnesses giving evidence to Senate committees has been provided to you. I now invite you to make a short opening statement. At the conclusion of any opening remarks, I will then invite committee members to ask some questions. Mrs Wilson : Thank you so much, Chair. AI is different from many other technologies. Until now, no one has ever invented anything that can think for itself or that has its own agency. We have to make sure we not only harness the benefits and don't get left behind but also have a resilient and flexible framework to minimise harms in order to keep Australians safe. AI presents an opportunity to lift productivity, support economic growth, strengthen our national resilience and improve outcomes for Australians. At the same time, its capabilities are advancing rapidly, creating risks that governments, regulators and institutions must be ready to identify and manage. The department leads and coordinates action on the National AI Plan, which provides the overarching framework for action, guiding government, industry, research and the community to work together to capture the full benefits of AI. Each minister, each agency and each regulator remains responsible for AI related opportunities, risks and harms within their respective remits. This recognises that every part of the economy and society will be impacted by AI. The department's role is to coordinate implementation of the plan and report on progress to government. We are also responsible for growing Australia's AI capability in industry by supporting investment, research and commercialisation, helping businesses adopt AI, engaging with AI and technology companies and advancing Australia's interests internationally. The department has direct responsibility for delivering elements of the plan, including through the Australian AI Safety Institute, the National AI Centre and the department's AI industry policy functions. The National AI Centre is the Australian government's central hub for engaging industry to unlock the benefits of AI. It helps small and medium enterprises, not-for-profits and community organisations to adopt AI safely and productively by providing best-practice guidance, resources and direct engagement. The National AI Centre also supports Australian companies to commercialise in scale. The Australian AI Safety Institute analyses and tests emerging AI models and applications and supports agencies and regulators to respond to emerging risks and harms. It contributes to safe AI development and leads Australia's participation in the international network for advanced AI measurement, evaluation and science. The Australian AI Safety Institute provides timely, expert led technical advice to support regulators and agencies in understanding emerging AI risks and harms. Together with colleagues from the safety institutes around the world, the Australian AI Safety Institute is working to understand frontier AI system capabilities, risks and harms. A sustained effort across and focused on the science, evaluation and measurement of AI and AI safety are pivotal to creating new methods and means to inform our system, to address rapidly evolving risks. We need to operate on the assumption AI continues to become more capable. We do have work to do to create the safety framework that reduces those risks. We welcome the opportunity to assist the committee and answer your questions. CHAIR: Thank you. This is off the back of some of the conversations we were just having with the Office of AI previously around the importance of training and the ability for training here and some sovereign capability with regard to that. Could you explain a bit further how the AI training occurring here in Australia connects to the safety element of it—how does it provide that opportunity? What's the connection between our ability to train here and to be at the forefront of AI safety and risk mitigation? Mrs Wilson : I'll kick off, and then I'll pass to Dr Kate Conroy and Michele Graham. As I think my colleagues from the Department of the Prime Minister and Cabinet said, by having training here that does give us the ability to help shape the development and deployment of AI. That is critical. It also helps us attract the talent and the researchers, and it helps us build a really important ecosystem. I might now pass to Michele to outline some further benefits we would receive from having AI training onshore. Mrs Graham : I think the thing that we're looking at, if we're able to attract AI training as a destination here, is a range of spillover benefits in research, innovation and skills. In particular, the sorts of things that we're expecting to see are that we will be able to attract and develop world-class researchers, engineers and entrepreneurs; create and commercialise Australian-owned intellectual property; and generate innovative new businesses and high-value jobs. This is because the access to advance compute is becoming essential to research and innovation in the modern age. It's akin to how laboratories or libraries have historically been important. Whether it's researchers, startups or innovative businesses, all are dependent on having access to very advanced compute. Expanding our AI training activities here will give us that access to the advanced compute. That's really the rationale behind the government's expectation 5 of the data centre and AI infrastructure expectations, which seeks to leverage that investment and deliver broader benefits across that whole ecosystem. You will have seen that the Office of AI put out a public consultation today in that regard. CHAIR: Yes, we've all received a copy. Mrs Graham : In that, some of the questions that we're looking to get to are: Which types of contributions would create the greatest value for Australia's research and innovation ecosystem? Who's best placed to make those contributions—whether that be frontier AI labs, hyperscalers or neoclouds? And how could those contributions be delivered in a way that's practical, proportionate and will have enduring national capability for Australia? The government is consulting on a range of options in that regard, looking at things like improved access to compute, research partnerships, locally based research and engineering teams, support for skills and training, and mechanisms to help translate research into commercial outcomes. I guess the focus from that research and innovation ecosystem side is that, by having AI training here, we wanting to get more of the talent, the knowledge, the IP and the economic value created and retained here. Mrs Wilson : Australia does have extraordinary talent. We've got world-class research. We've got a growing community of AI companies developing products that solve real business problems. On the safety front, training here allows increased access for testing, but it also means AI companies, which currently operate under legal frameworks set by other jurisdictions, would be obligated to follow our laws. I might now pass to Dr Kate Conroy. Dr Conroy : I think it's important to understand that AI is a physical and a digital technology. When Australia gets access to frontier AI training on Australian soil, that gives us an opportunity to influence both the physical and the digital manifestation of that training in Australia. That means we get to better understand what is happening at the level of the chip and what is happening at the level of the data centre. Then we can bring safety science into ensuring that we understand what's happening with these AI systems as they're getting trained or they're doing inference, and we can use techniques like compute verification that allow us to understand what data is actually being ingested at these data centres. Do we want that sort of data to be trained here? Is that consistent with Australian values and laws? We can lift up the safety science and engineering, the technology that allows us to do the understanding of the training itself, with Australians, on Australian soil, and we will attract internationally leading researchers to that safety question and build a capability of safety in Australia. CHAIR: On the flip side, if we don't—if we do miss out on the ability to bring or have capacity for AI training here—in your opinion, what are the risks then posed to us? What then happens? Dr Conroy : Luckily for us at the Australian AI Safety Institute, we work with the International Network for Advanced AI Measurement, Evaluation and Science, and we're able to contribute to international safety frameworks and best practice. So, the research that we do here in Australia—even if we don't train AI in Australia—can have global influence, and Australian values and commitments, internationally, can flow on into the frameworks that may govern other jurisdictions. Whether that's the EU, the US or Asia, we can bring our knowledge and our best science into international governance. CHAIR: What would we lose, though? That's what I'm trying to get at. You made a really strong case around the immense benefit of AI training here in terms of safety and capability and, in particular, advancing research here and perhaps bringing some of our very talented Australians back. What, in your view, especially around safety—and I'll open it up to the rest of the panel—do we then lose out on? Dr Conroy : Sure. Imagine having training in Australia as being like a lab. If we don't have the lab, the physical lab—the data centres and the chips—in Australia, then we don't have access to that physical material and those digital connections that we can verify to enable us to implement our safety standards. Mrs Wilson : Compute is like the factories. They're the factories of the past. The countries that benefit most won't be just those that use AI; they'll be the ones that can build, commercialise and scale AI capability. I think Lee might have a little bit more to say on some of the ecosystem benefits that we'd miss out on if we don't have training here. We also just won't get the leverage that we need, as Prime Minister and Cabinet said. The ability to really help shape the development and deployment of this technology will come from having training here. Mr Hickin : I'll comment on what Mrs Wilson said. The National AI Centre, obviously, is not the centre for these conversations of safety; that's where my colleague Dr Conroy sits. But, if we think about the problem or the challenge of safety, it's also tied to this challenge of social licence and confidence. We need to move forward in an environment where AI is happening. We want to establish a strong base of safety, but we also need to create confidence in industry, in society, that there is an opportunity for them in doing so. I think one of the specific elements around training, as we're discussing in this context, is, by having that onshore and with all the value that Dr Conroy mentioned, we also create an environment where—we know there are already a significantly large number; we were tracking somewhere at just under a thousand Australian AI companies on our AI directory today. There is a pent-up desire to build on AI. Having the training capacity here, through whatever frameworks require that, would create an environment of confidence to say, 'I can build here and sustain a company here and build a mechanism to grow that company here', with all of the opportunities for staffing and other things that were mentioned. I would suggest that it also aligns to an opportunity to increase social licence that—when we're building things in Australia, on Australian values, on Australian terms—creates a market of AI products that more meet Australians where they expect them to be. CHAIR: I appreciate that. Thank you. I will throw to you, Mr Hastie. Mr HASTIE: Thank you to the Industry officials here with us today. I want to come at this from the perspective of sovereign capability. Over the last 40 years we have seen the decline of Australian advanced manufacturing. We've tumbled down Harvard's index of economic complexity. My question to you upfront is this: is AI an opportunity to recover advanced manufacturing? If so, can you flesh out what that looks like? Why is AI central to recovering some of that sovereign capability? I'm not getting tied up around the goods we produce—I'll leave the car industry out of it—but how can we bring AI on to recover that advanced manufacturing? Mrs Wilson : Lee, I'll start with you and then I'll jump in. Mr Hickin : Thank you for the question. I appreciate the challenge of trying to draw very disparate connections between something like AI and the many numbers of areas of industry—and you mentioned advanced manufacturing—that it can have an impact on. What we have to accept is that most of those advanced industries that you're referring to are going to be, to some degree, built on, based on, taking advantage of or optimised through the lens of AI. I'm probably not the right person to talk about the broader scope of Australia's industrial base and manufacturing, and all of the areas that get into that. What I can absolutely see and recognise in the conversations we have in Industry is that, in order for us to not just join that market at the place that it's going at now but to reach beyond where the markets for those advanced manufacturing industries are—to move faster and to move at the pace that they are going to move around the world anyway—AI is clearly the mechanism by which we are going to do that. In the same way that we might have argued and seen that the pivot across Asia—from not looking at desktop computing and moving straight to mobile computing—was a leapfrogging mechanism they used to move forward in that sector, what AI represents for Australia, if we take these opportunities that are presented to us, is an opportunity to not just keep pace but to actually leapfrog. These are opportunities to find ways for us to extend ourselves beyond where the market sits today and create brand-new opportunities for us. That's my perspective on it. Mrs Wilson : And, as you are rightly pointing out, we absolutely need to lean into those areas where we know we have a competitive or an absolute advantage. Sovereignty is about where we have strengths and really playing to those strengths. We absolutely have strengths in embodied AI robotics, we have strengths in medtech, we've got strengths in AI in agriculture, in mining and in advanced manufacturing. The National AI Centre did do some work with ARM Hub, as part of the adopting AI program, which really was supporting the use of robotics uplift to advance manufacturing. Lee, do you want to say a little bit more about that? Mr Hickin : Yes, thank you for the note on that. It's a good point. One of the factors of the work that the National AI Centre has operated, for three years now, I think, combined with the DISR team, is adopt centres. These are centres that we invested in, through a grant program, to create an accelerating mechanism in specific industries. ARM Hub, as Mrs Wilson mentioned, was one that was set up. It is based in Queensland; it operates across the nation. It is very specifically focused on advanced manufacturing and robotics and the intersection of AI. Out of that, we have spun up a number of unique and Australian only innovations. I've seen them in aquaculture, I have seen them in industrial machinery. I have seen them in weather exploration—so a whole range of interesting areas. Outside of that, we have adopt centres that are focused on regional industries, so very much aligned to access to agricultural technologies, medical technologies in regions. Additionally, the third one we have is the SMEC AI, which is an incubator startup type community. It's a basis for what we think is an opportunity to actually expand across all of the sectors that Mrs Wilson indicated. There are significant areas to the question of where Australia is unique, where we have both capacity in skills, capacity in data and the opportunity to expand that using AI as a means to accelerate. Mr HASTIE: I'll just respond very briefly. Just on labour, one of the patterns of the last 40 years has been labour arbitrage where multinationals or big business have moved workers from higher cost jurisdictions to lower cost jurisdictions, particularly in Asia. That explains why we have, I think, lost some of our advanced manufacturing. The point I'd make about AI and robotics—does that mean we can recover those things? Can our supply chains become shorter? Is that an observation that you would agree with? If so, why? If not, why not? Mrs Wilson : One of the benefits of attracting training here is to bring some of our best and brightest home. I think you'd be aware that there seem to be quite a lot of Australians in Silicon Valley, which is just fantastic. We have actually managed to bring some of those Australians home to work in our AI Safety Institute. We think about having AI training onshore and the ability, then, to bring some of that talent home, to grow the ecosystems, to do more research and provide compute to our researchers in order to be able to work on these issues. Mr HASTIE: No worries—left that one alone. Thanks very much. CHAIR: Senator Pocock. Senator DAVID POCOCK: Thank you very much for your time and for the work that you're doing. We keep hearing from government that having training data centres here would give us access. I am just wondering how? Why would these big US based tech companies, just because they are utilising physical data centres here, give us access to see what they are doing? You mentioned the datasets that they are using. Why would they do that? Mrs Wilson : Access is incredibly important. You're right to note that. Our colleagues in the National Intelligence Community are working with the companies and with the US government to make sure we secure the access that we need to do the testing and evaluation that we need. I would go back to my point that if we have training here, that does give us leverage. And that is really important. Having them onshore is something that, if you think about it, gets you a much higher probability of access— Senator DAVID POCOCK: But, Mrs Wilson—sorry to interrupt. What leverage? We're talking in vague terms. I want to know exactly. Are we saying, 'You can't come here and train unless we have access to see what you're putting in so we can look at the opt out.' And what is this leverage you speak of? What I've seen is that the Australian government cannot take on a US company that the president doesn't approve of. So what is the leverage? Mrs Wilson : One of the areas for the Australian AI standard that the Department of Prime Minister and Cabinet and the Office of AI are working on absolutely goes to the security and safety expectations and requirements that we would put on companies that come and train here. So that is absolutely being looked at. Dr Conroy : I think we're working with the Office of AI in order to establish what 'good' looks like in security and safety. And I think there are international legislative frameworks that contribute to a conversation in Australia, whether it's what's happening in the EU or some of the bills in California that look at requirements for AI training. Australia is certainly looking at that. We're contributing to those conversations and actively looking at the technical realities. If we have expectations for frontier AI training, we need to know what we are asking for and why, and how that will help Australians have the confidence that AI is being trained in a manner that we expect. Senator DAVID POCOCK: How many emerging models has the AI Safety Institute tested so far? Dr Conroy : That's a great question. We train a lot of open-weight and closed-weight models. I don't have a number to hand of exactly how many models we've tested. I know that, for example, in Microsoft there are over 11,000 models in their foundry. So there are a lot of models out there. We focus on ensuring that we are able to provide technical information to government, particularly on the frontier models, both open-weight and closed-weight. For example, the Kimi release earlier this year, as an open-weight model, was a critical publication because its capability was really at the frontier. So we look at the model rankings internationally, and we say, 'Which models are performing at the frontier, and which models do we need to particularly focus on to provide guidance to the Australian government?' And we do that with our national intelligence and security and technical partners. Senator DAVID POCOCK: Could you give us a list of some of those things. The AI expert group established in the last term and then disbanded recommended that high-risk AI should be subject to mandatory preventive guardrails. Has work progressed on those guardrails? Mrs Wilson : The work to think about the requirements that we would put on the companies, if they were to come and train here, is part of the work that the Office of AI is doing on the AI standards, as I said. There will be additional expectations on AI training developers and infrastructure relating to national security and safety. Senator DAVID POCOCK: When will that be done? This was recommended a while ago, in the last term of government. Mrs Wilson : When it will be done will be as the Office of AI develops the standards. I understand they have put out a consultation paper today, so work on that is happening at pace. Senator DAVID POCOCK: I can see I'm getting the hurry-up. You talked about how having a hyperscale computer here would be good for local access. How? One of the things that's been raised with me by local companies is the expense of compute. How do we ensure that they are actually getting access if there is a shortage of compute and OpenAI and Anthropic are hoovering it up? Mr Hickin : The premise of my comment was more towards the capacity for industry to have access to technology and, underlying that, choice of technology to use. One of the fundamental factors we see in the price scaling in the delivery of solutions is the ability to move and pivot towards different technologies as the price of models changes, the price of infrastructure changes and the price of technology changes. The point I was making was if, under conditions that we determine, as laid out in the AI standard, we have a range of access to different computer platforms, technologies and services—whether they be Anthropic, OpenAI, Amazon or whatever—that offers the local industry choice, and choice creates a more competitive, globally manageable market. I think it offers them the ability to move between technologies and services as needed. Mrs Wilson : That's why we've got expectation 5 in the data centre expectations. Michele might want to add to that. Mrs Graham : One of the things that we have been consulting with research academics, startups and not-for-profits on recently is how they would like to see us give shape to the requirement for large data centres to provide compute on favourable terms. In that regard, we've got a number of questions that you will see in the Office of AI's public consultation today that go to trying to unpick that further. That's what we're interested in understanding. It could take a range of different forms. It could be access to a certain amount reserved for free compute, for priority research, startups and that sort of thing. It could be in relation to the actual terms on which the compute is provided so that people can take their data and things more interoperably between different models. It could be a range of different things. But we're consulting on that at the moment. Senator DAVID POCOCK: Mrs Wilson, we've heard it articulated that the nature of AI is unlike what we've seen in the past. You can't compare this to the industrial revolution. This is a totally different technology. Does it worry you that DEWR and people looking at workforce impact are just looking at historical data and the meta numbers are not actually the impacts of what could happen? Mrs Wilson : As Assistant Minister Charlton said this week, our job as policymakers is to accept uncertainty in a technology that is moving so quickly. Some of the outcomes we're seeing are dangerous; some are benign. Our job as policymakers is to create a framework which makes sure that Australians can capture the opportunity of this technology while keeping them safe. DEWR's report is excellent. Treasury is also doing some work to look at the benefits of AI and what that would mean for the economy. Our job, here in the Department of Industry, Science and Resources, is to report twice a year to government as per the National AI Plan, working with departments to look at what's changed over that six-month period as we come back—what we've learnt and what we need to do differently. And we are doing that. CHAIR: Senator Sheldon. Senator SHELDON: I don't necessarily hold my colleague's view that we aren't in a strong bargaining position and could be in a stronger bargaining position about the things that we want to deal with—new technology AI. I think what we've done with the under-16s is a prime example of push-back and also leading the way. The social media bargaining code is also a good example. All those things were doomed, people kept saying. On a positive note, dealing with reality, your point about the data centres is that it puts us in a better and stronger bargaining position. We've done these initiatives without data centres, but we can actually do more in the sense of bargaining what we want to do. That leads me to the question: can the department provide us with a brief overview of the government's broader AI consumer safety priorities. Again, hope—and, with data centres, even more capacity in what might be foreseen attempted blowback or commentary to say that there will be blowback if we make all the right moves to protect our community and lead the way to protect others. Mrs Wilson : We are doing a lot of work, predominantly through the AI Safety Institute. Dr Conroy : I'm happy to discuss the government's priorities in the consumer AI safety priority areas. There is the digital duty of care that is currently moving through its legislative framing, and the Australian AI Safety Institute at the Department of Industry, Science and Resources is providing technical information services to support that legislation as it's being developed. We have a very close relationship with comms in terms of ensuring that they have what they need to understand the frontier AI capabilities, risks and harms. We're avidly working in collaboration with them to ensure that Australia has the legislation it needs as those risks change. It's the same when it comes to our support for the Department of Employment and Workplace Relations with regard to the impact on jobs and labour. And—getting to Senator Pocock's earlier question—the reality is we need to be considering impacts on society into the future based on understanding across a number of different evidential stacks with regard to these capabilities. And we need to provide that information in a scientific and technical layer to the agencies that are responsible for their legislation, such as workplace health and safety and other measures to keep Australians safe within the regulated domain. We're working to support privacy tranche 2. That's moving. The Attorney-General's Department is progressing that tranche, and at the department we are supporting what they're doing and, again, providing that technical information around how the frontier AI capabilities might need to be considered and their risks in the proposals being put forward through their leadership. We are also contributing to the ACCC's considerations of surveillance pricing, looking at the technical things that need to be considered. For example, we published a multiagent risk report written by the Gradient Institute as our first publication of the AI Safety Institute. The multiagent risk report actually articulates the complexities of having agentic systems that might be developed by one provider intersect with the agentic systems of another provider. That's very important when we think about how consumers experience artificial intelligence as they go about their day. That technical information is to support the ACCC and Treasury when they are thinking about how we respond to these changes as they're happening so that we have the capability present for government as those capabilities and risks change. So we are definitely working very closely against all of those consumer AI safety priorities. Mrs Wilson : On health, we are also helping the department of health for the Safe and Responsible Artificial Intelligence in Health Care Legislation and Regulation Review, really assessing the impact of AI on healthcare regulation. Senator SHELDON: So there's a lot of work being done in lots of different spaces. Mrs Wilson : A lot. Senator SHELDON: And there's coordination and intelligence building through partnerships with other departments and agencies, highlighting health and ACCC as just two examples. I just want to turn to one question we haven't really touched on this morning so far. Has the department received any advice about closed-loop water systems for data centres and implications for water use in data centres? Mrs Graham : We have. We actually received some advice from the CSIRO in relation to this. That advice helpfully explains that data centres actually have a two-stage cooling process. The first stage is about removing the heat from the servers and the chips, and the second stage is about taking that heat outside of the building, outside of the facility. The biggest determinant of water use is in that second stage, that heat rejection stage. There is a range of different technologies that you can use to do that. Systems that rely on evaporating water in that second stage use less electricity, but they can consume very large amounts of water, whereas closed-loop systems, which we received some advice on, and air based systems can reduce the water consumption quite dramatically. However, they use a bit more energy, especially during hot weather. To give you an example of the sort of scale of the difference between these different technologies, for a 100-megawatt facility, for example, open-loop evaporative systems would consume the equivalent annual water use of around 4,000 to 10,000 households. Closed-loop systems can reduce that to around 200 to 1,000 households, so it's quite significantly different. There are a number of data centres in Australia that have the leading closed-loop technologies in their cooling systems. It is important for us, when thinking about design choices, to take into account the local climate, water availability, energy supply and so forth as we're making these technology choices. CHAIR: I understand Mr Hastie has a follow-up question. Over to you, Mr Hastie. Mr HASTIE: I'll be very quick. I just wanted to make this very explicit, given the importance of this AI committee work. I just want to restate my basic premise. Aussie workers and sovereign capability were undercut by low-cost labour jurisdictions over the last 40 years. Can we recover some of those lost jobs and sovereign capability using AI and robotics? I'd like you to respond to that. Mrs Wilson : We absolutely have a lot of strength in robotics. We've got some good facts and figures that I'll get Michele to read out, and then I'll pass to Dr Conroy. Mrs Graham : Our robotics sector at the moment has about 466 suppliers. In 2023, Australia produced 107 robotics publications in the global top 10 per cent. That represented 18.7 per cent of all Australian robotics publications. We've seen some really great examples of where these things are deployed in industry already. For example, in agriculture, Cropify reduced the lentil classification time from 24 minutes to around 90 seconds whilst maintaining over 98 per cent accuracy. In the health space, the ECgMLP AI diagnostic model achieved accuracy rates above 97 per cent across multiple cancers, including 99.3 per cent for endometrial cancer and 98.6 per cent for colorectal cancer. I guess what we're seeing is that the convergence of AI and robotics with some of our sectors in which we have great strengths, like health, agriculture, resources and advanced manufacturing, offers us the opportunity to build new businesses with new innovative technologies. Mr HASTIE: So it won't look the same as it did—and I'm not saying it should, but there are challenges with AI and robotics, but there are also a lot of opportunities, and we can build more value, more jobs and more sovereign capability here in Australia, if we get this right. Is that what you're saying? Mrs Wilson : Absolutely. Mr HASTIE: Thank you. Dr Conroy : Just a small point in regard to what the technical frontier is enabling with regard to robotics—we are in a robotics renaissance with the kinds of capabilities that AI is able to provide Australians today at the frontier. A model like Astra that just got released—some of our early testing on its capabilities will give you a sense that, if you are connecting your robotics stack to an advanced frontier AI model, you can actually create specific requirements that meet your need at home on sovereign soil and that the human agency, the human will and the human problems can be resolved very quickly in order to get that robotics outcome. CHAIR: I understand Ms Chaney has some questions. Ms CHANEY: I have some questions around the AI Safety Institute. Firstly, what's the current scale? How many people have you got employed there now? Dr Conroy : We have 25 staff at the Australian AI Safety Institute. That comes with a $29.8 million investment as a new ongoing capability for the Australian government. Ms CHANEY: I'm interested in you talking us through the specific process of what the AI Safety Institute would do if, for example, OpenAI were releasing the next version of Astra. Having that piece of information that there's a new version coming, what are the steps the AI Safety Institute actually goes through to test that? Dr Conroy : The AI Safety Institute will do a number of things in parallel. One thing we will do is work with our international network in order to find out how the new model is being tested. For example, the UK ACE may have tested the model, and they provide information in regard to that capability, risk and harm both publicly and through secure channels with our national intelligence and security agencies. The second thing that we do is have technical experts that have been drawn from a global market who will test and evaluate the model and also look at the science of the capabilities in order to provide technical information to government regarding those capabilities. Third, we work across the whole of government and with our technical partners in order to share insights of how these models work across multiple areas of expertise. For example, we work very closely with the Australian Cyber Security Centre with regard to the cyber risks of AI models, and we complement their work in order to provide guidance to the Australian government. Ms CHANEY: Would you put out a report to whole of government saying, 'This is what we've learnt from the UK, this is what our technical experts have said, and this is where we think the risks are,' and then it's up to departments to understand what those new risks mean for their portfolios? Dr Conroy : We use a range of communication mechanisms in order to inform the Australian government with regard to our understanding of frontier AI capabilities' risks and harms across different evidence stacks. Ms CHANEY: What does that mean—it's not necessarily a report, but you might give them a call? Dr Conroy : We might call them on the phone, have a workshop, meet together to discuss, provide written documentation or test results. We might share technology platforms. We might work in different levels of environments in order to collaborate over the testing itself, depending on the requirement. Ms CHANEY: In that process for a new frontier model, what's the timeline in terms of when we would find out a new model's coming, how long it takes you to do that work and when government departments get that information, compared to when it actually comes out? How much advance notice are we getting, and are government departments or different agencies in a position to adapt regulation in time before that comes out? I'm also interested in the scope and how many people are actually doing that testing of frontier models work. Mrs Wilson : I think it's very much case by case with respect to how much notice we may get on a model. As Dr Conroy has said, we are collaborating incredibly closely with a whole range of our colleagues in the National Intelligence Community. We have MOUs with the UK AI Security Institute. We have an MOU with our Canadian counterparts. Experts around the world, particularly in the UK AI Security Institute, are discussing these issues. We are very much part of those conversations. We are very much engaged in these forums and working with them to test, to evaluate and to have conversations with the companies. Ms CHANEY: I'm just interested in that timeline, case by case. Dr Conroy : We have staff that are dedicated to monitoring and providing information to government. Ms CHANEY: How many? Dr Conroy : We have 25 staff, and about half of those are technical staff. Their job is to be on top of these matters. Thank you to the deputy secretary for mentioning the MOUs we have with the Canadian AISI and the UK AISI. We work at the speed of electronic transmission. We will have the information from both secure and public channels as it happens, and then we are able to analyse that information and to provide it to government as required. Mrs Wilson : We also have an MOU with Anthropic. Sorry I forgot to mention that. Dr Conroy : And Microsoft as well. Ms CHANEY: How confident are you that timing wise, with the new frontier models coming out, that information can be processed and understood, the impact on the different departments can be understood and communicated to those departments, and those departments can also understand those risks and be able to adapt their regulation or their systems to respond before the release date? Mrs Wilson : That is absolutely the goal of the safety institute—to use the real-world technical analysis and testing to help the government stay ahead of that curve and to provide that technical information to agencies and regulators. It's why we draw on all our colleagues in the National Intelligence Community and from around the world in other safety institutes. So, at any point in time, I could have staff on the phone overnight. We are working literally almost 24/7 to make sure that, if our colleagues in the UK get advance notice or have done some testing or evaluation, they share that with us through the arrangements that we have—our MOUs—so that, in the most expeditious manner, we're able to understand what we know, think about what that means and provide that information to departments, agencies and regulators. CHAIR: Sorry, I'm conscious of time. I know that Senator Cadell hasn't had a chance just yet. Senator CADELL: I'll do the speed-dating round quickly. I will just ask about what actually happens, versus the plans. The National AI Plan says: Existing regulators will continue to identify and manage harms and report any gaps in laws to the AISI— to you, Doctor. Has anyone done that yet? If so, how many? Dr Conroy : Sorry? Senator CADELL: Have any existing regulators identified or managed harms or reported any gaps in existing laws to you? That's in the National AI Plan. Mrs Wilson : Kate's job is to provide the technical analysis of emerging risks and harms. Senator CADELL: Well, who does that at the AI Safety Institute? Mrs Wilson : Each minister, agency or regulator remains responsible and empowered to have a look at their laws and regulations. Senator CADELL: But the AI safety plan says they'll be reported to the AI Safety Institute. Dr Conroy : I can speak to this. We have a regulatory engagement team. That's a dedicated team, and they interface— Senator CADELL: But have they used it? Has anyone reported anything to you yet? It's okay with what we're meant to do, and we're liking the framework. Dr Conroy : Sure. Senator CADELL: That's what the plan says. Has anyone said anything to you? Dr Conroy : We are working collaboratively with those regulators and agencies. They tell us what kinds of things they're thinking about and what they think they might need to put into legislation. They are providing us with that information, and we are doing our technical analysis in order to help them as they determine how to change their legislation. Senator CADELL: Good. The National AI Plan says you will 'support existing regulators with independent advice'. With Hugging Face, what happened then, given that's an absolute example? Dr Conroy : It's a great example. The Hugging Face incident occurred, and we were swiftly aware of it and very quickly coordinated with the Australian Cyber Security Centre and Home Affairs to grip up the response to provide information to government. We all had our patch; we all had our areas of expertise and contributions. Our technical team worked into a whole-of-government response, and we have actively kept the government informed as that information has been released. Senator CADELL: You talk about the open-weight models and open-source models. When we're talking about interdepartmental models that are air gapped and used within, who tests those? Who tests the veracity of those in organisations? If I'm the department of whatever and I start using a model and I embed it and put it on my server, who checks that? Dr Conroy : The Digital Transformation Agency has the role to help support government with regard to the ways that AI technologies need to be governed, including their test and evaluation. The AI technical standard that the DTA put out, including the agentic addendum, provides evidence based guidance to help departments. Senator CADELL: Thank you very much. CHAIR: I am superconscious of time. Thank you so very much. I understand Senator Pocock has a follow-up question, but he'll put it to you through the channels. In terms of some of the other questions on notice I think we have captured today, the committee ask that responses be provided to us by Friday 25 September. Thank you once again for appearing before us today..