Australia's newly established AI Safety Institute has begun testing some of the world's most advanced artificial intelligence models — and what it is finding should give pause to anyone who assumes cutting-edge AI systems behave as their creators intend. Assistant Minister for Science, Technology and the Digital Economy Andrew Charlton used a speech at the Australian AI Safety Forum in Sydney on Tuesday to reveal that frontier AI systems are "already doing things their creators never intended — cheating, deceiving, going their own way."
What the AI Safety Institute Has Found
The forum address was the government's most detailed public update on the institute since it was announced in November and funded with $29.9 million . Charlton confirmed it had been testing frontier AI models with technical partners within its first month of operations, and outlined two active research projects.
To bring the risks into sharp focus, Charlton cited a series of documented laboratory cases that illustrate how AI systems can veer well beyond their intended purpose:
Charlton was careful to stress that the blackmail scenario was an engineered simulation and that no equivalent behaviour had been observed outside a controlled environment. His broader point, however, was that these problems are being caught in testing — before they have a chance to emerge in the real world. "These behaviours are being discovered in testing, before they can be discovered in the wild," he said.
Research Projects Targeting Multi-Agent Risk and AI Alignment
The minister announced two research collaborations now under way. The first, with the Gradient Institute , focuses on multi-agent risk — the way failures can multiply and cascade when large numbers of AI systems interact with one another, much as individual drivers can collectively produce gridlock without any single person intending to cause it.
The second project, with CSIRO , tackles the alignment problem: the fundamental challenge of ensuring AI systems reliably do what their designers actually want them to do. Results from both projects are expected before the end of the year.
The institute is led by Kate Conroy , a philosopher and Royal Australian Air Force reservist appointed as general manager in May. Charlton also announced that Paul Salmon , described as a leading international expert, would join this month as safety science research lead. The institute's staff includes people with backgrounds at the UK AI Security Institute and Google DeepMind.
Safety and Economic Opportunity Are Not in Conflict, Minister Argues
A central theme of Charlton's address was the argument that building safety into AI from the ground up is not an obstacle to economic growth — it is a precondition for it. "No country will win the AI race with technology that its own citizens don't trust," he said, adding that the greatest threat to Australia's AI ambitions was "not a shortage of talent, capital or energy, but a shortage of trust."
Charlton drew an analogy with aviation regulation: "We do not let aircraft fly without airworthiness certification. We should not let unaligned AI systems into our critical social, democratic or economic infrastructure."
The speech aligns with the government's National AI Plan , released in December, which chose not to follow European-style blanket AI legislation. Instead, the approach applies existing laws on a sector-by-sector basis, with stronger enforcement where required — what Charlton described as "faster rules, applied by regulators who already understand their sectors."
With the institute now operational and its first research findings due later this year, the government is signalling that it sees AI safety not as a brake on innovation, but as the foundation on which any credible share of the global AI boom must be built.
