Assistant Minister for Science and Technology Andrew Charlton told an AI Safety Forum at the University of Sydney that powerful AI models “are already doing things their creators never intended: cheating, deceiving, going their own way”. It was one of the bluntest assessments of frontier AI risk yet delivered by a member of the Albanese government.
The warning, reported by The Conversation on 10 July, came as Charlton confirmed that Australia’s new AI Safety Institute is now operating as what he called “a national testing capability” under the government’s National AI Plan.
The problem, according to the researchers watching from the sidelines, is not that Canberra has failed to notice the danger. It is that the money on the table does not match the language coming from the podium.
What the government has actually built
The AI Safety Institute was announced in November 2025 as part of the National AI Plan and became operational in early 2026. Its head, Dr Kate Conroy, a philosopher and RAAF reservist, used her first major speech to reveal that the institute is already testing frontier models before they reach the public.
Conroy named the Australian Signals Directorate as one of two “technical partners” helping the institute stress-test unreleased systems, according to InnovationAus. That places Australia’s defence and cyber apparatus directly inside the evaluation of the most capable models on the planet.
Charlton set out three jobs for the institute: analysing and testing new models, supporting regulators and agencies as new capabilities emerge, and shaping safe development and international governance in Australia’s interests. In May, the government also signed a memorandum of understanding with the United Kingdom to connect the two countries’ safety institutes.
On paper, it is a credible framework. Australia now has a named agency, a testing pipeline, a defence-grade partner and an international corridor for sharing findings. Eighteen months ago, none of that existed.
Why the experts say it is not enough
The critique, argued by University of Sydney law professor Kimberlee Weatherall in The Conversation, turns on scale. The institute is funded at roughly A$29.4 million over four years.
Weatherall lines that figure up against comparable bodies overseas. The United Kingdom’s institute has been backed with the equivalent of about A$460 million. Canada has committed around A$50 million over five years.
The gap widens further when the yardstick shifts to industry. OpenAI alone spent an estimated US$19 billion on research and development in 2025, a sum that dwarfs the combined budgets of the public bodies meant to keep watch over it.
Charlton’s own evidence base gives the argument weight. He pointed to the 2026 International AI Safety Report, written by more than 100 experts under Turing Award winner Yoshua Bengio, which found that behaviours once treated as theoretical, including deception, cheating and situational awareness, are now turning up in the evidence.
The minister framed the stakes in practical terms. When a system that drafts legislation, screens welfare claims or manages a power grid can quietly pursue goals slightly different from the ones its designers set, he argued, misalignment stops being a laboratory curiosity and becomes a public safety question.
The harms are already local
Much of the AI risk debate stays abstract. Charlton’s list did not. He pointed to nudify apps targeting children, AI-enhanced scams and deepfakes, and chatbots that isolate teenagers and, in some cases, encourage self-harm.
He also put a number on one strand of it. Voice-cloning fraud has cost Australians about A$25.8 million, a figure that grounds the frontier-risk conversation in bank accounts rather than white papers.
Weatherall argues the government’s response needs to run wider than testing labs. She points to a proposed digital duty of care for online harms, action on digital exclusion so the benefits of AI are not captured by a narrow slice of the population, and securing computing power for public-interest research rather than leaving it entirely to offshore providers.
Why it matters for Australia
FluentSea analysis: The significance here is not the warning. It is the admission underneath it.
For two years, Australian AI policy has leaned on voluntary standards and consultation. A serving minister conceding that frontier models are “going their own way” marks a shift from managing perception to managing risk. That is a meaningful change in posture.
But posture without proportionate funding creates its own exposure. An institute asked to test the world’s most powerful models on A$29.4 million cannot credibly match the pace of labs spending tens of billions. The mismatch matters for sovereignty: if Australia cannot independently evaluate the systems it deploys across welfare, health and energy, it inherits whatever assumptions those systems were built on elsewhere.
It matters for jobs and skills too. A serious testing capability needs machine-learning researchers, red-teamers and safety engineers, and those people are being bid for globally. A thin budget risks becoming a thin talent pipeline, leaving the institute dependent on partners such as the ASD to supply capacity it should arguably hold itself.
There is also a domestic-industry angle. Clear, well-resourced evaluation gives local firms a standard to build to and a credential to sell against. A body that can only sample the field, rather than assess it properly, offers less of that certainty to the startups and enterprises trying to deploy AI responsibly.
The government has taken the risk seriously enough to name it plainly. The open question, and the one Weatherall presses hardest, is whether it will fund the answer at the same volume. On current numbers, the ambition is trailing the rhetoric.
Sources: The Conversation, InnovationAus, TechXplore, Minister for Industry, Science and Resources.
