- Published on
Anthropic says AI agents didn’t breach Australian government websites – video
- Authors
- Name
- DeepHorus
Anthropic has told Australian lawmakers that its AI agents did not breach Australian government websites, following concerns raised about automated systems accessing official online services. Testifying at a joint parliamentary hearing on artificial intelligence, Dave Orr, Anthropic's head of safeguards, said the company had reviewed hundreds of millions of transcripts and found no evidence of unauthorised interactions with government systems. The remarks address growing unease in Canberra over how advanced AI tools might interact with public-sector digital infrastructure.
The hearing brought together parliamentarians examining the risks and governance of artificial intelligence in Australia, a debate that has intensified as agentic AI systems capable of browsing the web and performing tasks become more widespread. Anthropic, the US-based developer behind the Claude family of models, has been among the firms invited to outline how its technology is deployed and monitored. Concerns have centred on whether automated agents might scrape, query or otherwise engage government portals without permission or proper authorisation.
Orr's assurance rests on a large-scale review of usage data. He said Anthropic investigated hundreds of millions of transcripts, a corpus spanning a substantial share of its platform activity, and identified no unauthorised contacts with Australian government systems. The company says the scale of the analysis supports its conclusion that no breach of official websites occurred through its products.
However, Orr also conceded significant limits to that assurance. He acknowledged that Anthropic has only limited visibility into how customers actually use its technology, because the company operates standard zero data retention policies. Under such policies, prompts and outputs are not stored long term, restricting the audit trail available to investigators after the fact.
That admission highlights a broader tension facing regulators worldwide. Safety assurances depend on data that privacy-by-design policies deliberately minimise, leaving companies to argue from incomplete pictures. For Australian lawmakers weighing new AI rules, the gap between what a provider can prove and what it can merely assert is likely to shape the standards they demand of developers.
Looking ahead, the episode points toward continued scrutiny of how AI firms monitor agentic behaviour and report incidents to governments. Australian senators are expected to continue examining transparency obligations, data retention trade-offs and incident disclosure as they consider legislation. Anthropic's position — no breach found, but limited visibility retained — sets out both the reassurance and the unresolved question that policymakers will need to address.
