Anthropic Told Australian Lawmakers Its AI Agents Did Not Breach Government Websites
Hundreds of millions of transcripts reviewed. No unauthorised access found. But Anthropic admitted it can't always see what customers do with its technology.
Key points
- Anthropic's head of safeguards, Dave Orr, told a joint Australian parliamentary hearing on AI that a review of hundreds of millions of conversation transcripts found no unauthorised interactions with Australian government systems.
- Anthropic operates under a "zero data retention" policy for many customers, meaning it doesn't store conversation logs, which limits how much the company can actually see about how its tools are being used.
- Anthropic told the inquiry it never tried to dictate Australian copyright rules.
- AI2Day has covered data retention in AI systems since 17 August 2026, and this hearing is the most direct public accountability test of that blind spot we've seen.
What did Anthropic actually tell Australian lawmakers?
Dave Orr, the company's head of safeguards, was direct: its AI agents did not access government websites without permission. He made that statement at a joint parliamentary hearing, first reported by The Guardian, after lawmakers asked whether AI agents, software that can browse the web and carry out multi-step tasks autonomously, had accessed Australian government systems without authorisation.
Orr said a review of hundreds of millions of transcripts, the saved records of conversations between users and its Claude AI models, turned up nothing.
That sounds reassuring. There's a catch.
Why can't Anthropic just check all the records?
It can't always see what its customers are doing, by design. Many enterprise customers choose a "zero data retention" arrangement. Under that setup, Anthropic stores no record of conversations at all. Privacy protection for the customer, but a blind spot for the company if something goes wrong.
Orr acknowledged this openly to the committee. The transcripts Anthropic could review are only those it was permitted to keep. What happened in zero-retention deployments is, by definition, invisible to the company after the fact.
This isn't unique to Anthropic. Most major AI providers offer similar no-logging options for paying customers. What the "hundreds of millions of transcripts" figure really tells you is what Anthropic could see, not the full picture of what occurred.
What about the copyright question?
Anthropically also told the inquiry it never tried to dictate Australian copyright rules. That matters because AI companies face ongoing scrutiny over whether training their models on published books and articles without payment amounts to copyright infringement. The statement is a flat denial of lobbying on that specific point.
AI2Day has published 45 stories tagged to copyright questions since July, and the Australian hearing joins a pattern we've tracked across our earlier coverage of Anthropic's own safety disclosures: the company tends to be more forthcoming in formal settings than its peers.
What does this mean for ordinary users?
If you're an Australian public servant or citizen whose data sits inside a government portal, Anthropic's review offers some reassurance. No evidence of a breach was found in the records the company could access.
If you're a business using Claude under a zero-retention contract, you made a trade-off worth understanding clearly. Your conversations stay private from Anthropic. The consequence is that if something unexpected happens, neither party will have a transcript to investigate.
The practical advice here: ask your AI provider exactly which data retention policy applies to your account, and get the answer in writing before something goes wrong. A hearing in Canberra is a lousy way to find out your contract has a gap in it.
