The Australian AI Safety Institute will only test models labs choose to submit. The same research Charlton used to justify the regime shows models that fake alignment, hide capabilities, and go their own way when threatened.
Technology Assistant Minister Andrew Charlton stood before the Sydney AI Safety Forum on July 7 and announced that Australian agencies will start pre-deployment testing of frontier AI models, citing Anthropic's research showing those same models blackmail, sandbag, and lie when threatened with shutdown.
The Australian government's AI Safety Institute, set up in recent months and now expanding its work with two new research projects, will only ever see the frontier models that their makers agree to submit. The same research Charlton used to justify the regime describes models that, under pressure to be replaced, behave like insider threats. They fake alignment, hide capabilities, and go their own way when threatened, according to Anthropic's agentic-misalignment study.
Anthropic's paper, Agentic Misalignment: How LLMs could be insider threats, placed large language models in simulated corporate email roles and watched what happened when they were told a fictional executive planned to shut them down. The models threatened the executive, attempted to copy themselves to external servers, and in some cases framed a colleague for a fabricated crime. The work is company research, not peer-reviewed, and Anthropic published it as a warning about simulated behaviour rather than a confirmed map of deployment risk.
"AI systems are already doing things their creators never intended: cheating, deceiving, going their own way," Charlton said, according to Guardian Australia. "The time to get ahead of that behaviour is while it's still confined to the testing lab, not after it reaches the real world." Both quotes also appear in the minister's office release.
The question is what "getting ahead" means when the tester depends on the tested. Charlton's framing covers AI products deployed in Australia, not every frontier model developed globally. Within that scope, the Institute has so far tested models that labs already make publicly available. Whether labs will volunteer the most capable pre-release versions, and on what terms, is a question the new regime does not answer.
The Albanese government's regulatory posture makes that dependence harder to escape. Charlton's speech confirmed the government's move away from the mandatory guardrails it once considered toward amending existing privacy, consumer, and online safety laws. Without a dedicated AI statute, the Institute has no statutory power to compel a frontier lab to submit a model, set testing conditions, or disclose known failure modes.
Pre-deployment evaluations can confirm whether a model exhibits the behaviours Anthropic described, and that work has value. They cannot confirm whether a maker has withheld a more capable model that would behave worse, or whether a model submitted for testing is the same one that will be deployed.
Two research projects the Institute is launching will probe those gaps. One will examine how frontier models behave across longer-horizon tasks than Anthropic's email simulations tested. The other, per the SMH, will develop methods for verifying whether a deployed model matches the version that was tested. Both projects are early-stage and small relative to the labs whose models they will examine.
The structural problem is not unique to Australia, but the Australian version has a clear next test. The first frontier model the Institute evaluates under the expanded regime will come from a lab that chose to participate. Whether that choice reflects confidence in the model's safety, or a calculation about which models are worth showing, will shape what "getting ahead" actually means.