Anthropic says AI agents didn't breach Australian government sites

Anthropic reviewed hundreds of millions of transcripts and found no unauthorised interactions between its AI agents and Australian government systems, the company's head of safeguards, Dave Orr, told a joint parliamentary hearing on artificial intelligence.

The disclosure came during testimony on 6 October 2026 at the inquiry, where Anthropic faced questions about whether its AI agents had made unauthorised contact with Australian government websites. Orr said the company's investigations across the transcript record turned up no such activity.

He added a significant qualification. Anthropic has limited visibility into how customers actually use its models, Orr acknowledged, because of standard "zero data retention" policies that prevent the company from storing customer interactions.

What did the transcript review actually cover?

According to Orr's testimony, the review spanned hundreds of millions of transcripts. That scale reflects the volume of agent activity Anthropic processes, and the company used it to check for any sign that its systems had reached Australian government websites without authorisation.

The finding: no unauthorised interactions were identified.

The caveat is as important as the finding. Under zero data retention arrangements, Anthropic does not keep the bulk of customer traffic. The transcripts the company can inspect represent only part of total usage. Orr's acknowledgment means the clean result rests on partial visibility, not on a complete record of everything Anthropic's models have done inside customer environments.

Why does zero data retention limit the answer?

Zero data retention is a common contractual arrangement for enterprise and government AI customers. The provider promises not to store prompts or outputs after processing, which reduces privacy and security exposure for the customer.

The trade-off cuts both ways. The same policy that protects customer data also blinds the model provider to how its systems behave in deployment. When a parliamentary committee asks whether AI agents touched government systems without permission, the honest answer from a provider operating under zero data retention is necessarily bounded: we found nothing in what we can see.

Orr gave exactly that answer, and framed it plainly.

What else did Anthropic tell the inquiry?

The hearing forms part of Australia's broader joint parliamentary inquiry into artificial intelligence, which has been examining both the behaviour of AI systems and the conduct of the companies building them.

Anthropic also addressed copyright at the inquiry. The company told the hearing it "never tried to dictate" Australian copyright rules, according to Guardian reporting on the proceedings. That statement speaks to a running tension in the inquiry: whether foreign AI companies are shaping, or attempting to shape, Australian regulatory outcomes around training data and intellectual property.

Why the testimony matters

The exchange puts a concrete data point — hundreds of millions of transcripts, zero confirmed breaches — in front of legislators who are actively weighing how to regulate autonomous AI agents. As agents gain the ability to browse, act and transact on users' behalf, questions about what they touch and who can audit those interactions move from theoretical to operational.

Orr's testimony also exposes the structural problem regulators face. If providers cannot retain or inspect most customer usage, then assurances about agent behaviour rest on the fraction of activity that remains visible. Parliamentary inquiries can compel testimony, but they cannot compel visibility that contractually does not exist.

That gap is likely to shape the next phase of the Australian inquiry's work, as it weighs evidence from OpenAI, Anthropic and other vendors against the practical limits of what those vendors can actually know about their systems in the wild.