Anthropic Tells Parliament: No Breach of Government Systems, With Caveats
Anthropic has told federal parliament its models show no unauthorised contact with Australian government systems.
Dave Orr, the company's head of safeguards, gave the assurance to the joint parliamentary inquiry into artificial intelligence. He said reviews of hundreds of millions of transcripts found no unauthorised interactions The Guardian. On the company's account, no breach. Nothing to disclose.
Orr said Anthropic has limited visibility of customer use because of standard 'zero data retention' policies. If the prompt is not stored, it cannot be searched later. That limit shapes any later check for misuse.
Orr also said whether a firm reported a breach of government data would most likely turn on internal policy, not the law Bellingham Herald. Anthropic said it would not have waited to tell the government if one of its models had hacked a public site Yahoo News Australia. That was put as intent. There is no confirmed hacked site in this evidence.
Copyright and propaganda claims
On copyright, Anthropic told the inquiry it is not seeking to dictate the law on model training The Guardian. OpenAI and Anthropic had earlier urged Australia to rethink a ban on using Australian creative content to train AI models Reuters. Australia refused to let AI firms bypass copyright law for training Reuters. The line put to parliament is that parliament sets the rule.
Anthropic separately said foreign actors have used 'data poisoning' to corrupt models with government propaganda The Nightly. It did not give numbers or name the governments involved. The term covers inserting misleading material into training or retrieval data to shift what a model outputs, a bit like slipping dud entries into a reference book.
How the inquiry fits together
OpenAI and Anthropic chief executives were called to appear at the Australian AI inquiry Reuters. Anthropic later advised it would not attend the Senate committee hearing on AI on 1 October Reuters. Orr's evidence came at the later joint hearing on government systems and safety. The Senate inquiry is examining possible impacts of AI and data centres on communities, industry, water and energy.
Anthropic said in March it would sign an agreement to share its economic index data with the Australian government Reuters. That was described as a future agreement.
The broader context here is that Senate timetables are often untidy, and this one is no exception. A transcript review can only clear what the company can still see, and zero retention narrows that view by design. Policy-led reporting leaves the Commonwealth relying on goodwill unless a specific legal duty is triggered. The copyright tone has softened, but the commercial interest remains.
In my view, parliament was given a clean result from a bounded search, a frank caveat about visibility, and a reminder that disclosure rules are not settled. For members drafting recommendations, those boundaries are the substance.
