Anthropic to Let Australia’s AI Safety Institute Test Its Models Independently

Australia’s AI Safety Institute will soon be able to test Anthropic’s frontier models on its own terms, the company says. In remarks prepared for Tuesday’s parliamentary inquiry into AI in Sydney, Anthropic general counsel Jeff Bleich wrote that the company is “finalizing an agreement with the AI Safety Institute to enable it to run its own independent evaluations of our models,” Forbes Australia reported.

That goes further than the memorandum of understanding Anthropic signed with the federal government in April, which spoke of joint evaluations and sharing findings from the company’s own testing. The institute was set up last year with $29.9 million in federal funding. Bleich appeared alongside three colleagues, including head of safeguards David Orr and ANZ head of policy David Masters.

The day belonged mostly to OpenAI’s apology over its Medicare breach response, and Anthropic leaned on the contrast. “We have not found any cases where it interacted with Australian government systems in some sort of unauthorised way,” Orr told the committee, adding that Anthropic would notify the government “within a matter of days or sooner” if it detected anything. Prime Minister Anthony Albanese has said it took OpenAI 84 days to tell the government about the Medicare incident. Information Age reported that the government is weighing mandatory reporting of AI-related cyber incidents, an idea OpenAI said it would welcome.

What it means in practice

Independent access would give the institute something it has lacked so far: the ability to check claims about the most capable models itself, rather than relying on what the labs choose to share. The terms of the agreement, including which models are covered and whether any results will be published, have not been released.

Anthropic also used the hearing to press its copyright case. Bleich said the company respects Australia’s rejection of a broad text and data mining exception but is discussing an alternative mechanism with the government. Its September submission floated an approach in which creators could opt out, and acknowledged that AI companies may be pushed to pay. That debate, explored in our report on the opt-out copyright option, is bound up with Anthropic’s interest in investing up to $15 billion in Australian data centres, including its first lease for a Queensland AI campus.

What to watch: the signed agreement and its scope, whether OpenAI and other labs offer the institute the same access, and whether the committee recommends mandatory incident reporting in its final report.