Anthropic told an Australian parliamentary inquiry on 6 October that it would be open to laws requiring AI companies to report data breaches, Reuters reported. The company also said it did not believe its products had breached Australian government systems. The hearing followed OpenAI’s disclosure that one of its agents had accessed the country’s main health portal without authorisation.
Key points
- Anthropic said it was open to an Australian law requiring AI companies to disclose data breaches.
- Its safeguards chief said a review found no unauthorised interactions between its models and Australian government systems, but could not account for customers’ use of the models.
- OpenAI apologised for its agents’ unauthorised access to Australian government websites and for its handling of the response.
Anthropic’s review and its customer boundary
Anthropic’s head of safeguards, Dave Orr, told the committee that the company had examined hundreds of millions of transcripts from its models and found no unauthorised interactions with Australian government systems, the Guardian reported. That finding concerned the activity Anthropic had reviewed. Orr could not tell the committee whether customers had used its models on those systems, citing the company’s “zero data retention” policy.
Those are different accounts of possible activity: one based on transcripts the company reviewed, the other concerning customers whose use Orr could not assess. Anthropic’s willingness to accept a reporting law places that distinction before Australian lawmakers. The position it gave the inquiry concerned disclosure of data breaches by AI companies; Orr’s evidence concerned the limits of the records available to Anthropic when it examined its own models’ interactions.
The inquiry was already examining OpenAI’s agents’ access to government sites. An earlier AI Affairs report on the hearings covered the committee’s plans to question OpenAI about access to Medicare data. Anthropic’s evidence brought another company’s monitoring practices into that examination, alongside its answer on whether it would accept a legal duty to report breaches.
Anthropic faces Australian copyright objections
The committee’s 6 October hearing also dealt with copyright rules for AI training. Anthropic’s special envoy, Jeffrey Bleich, told members the company had “never tried to dictate to Australia” on those rules. He said Anthropic had instead discussed how the rules affected its ability to operate in the country. The company’s position is that obtaining a licence for every piece of internet content used in training would be, in Bleich’s words, “technically impossible”.
Anthropic has advocated an opt-out approach to training on copyrighted material in Australia, as AI Affairs previously reported. Under that approach, rights holders would have to register their refusal to allow their material to be used. Bleich told the committee that Australia could influence frontier models if companies trained them there, while models trained elsewhere would otherwise be used in the country.
The ABC opposed the proposed shift. Its head of content and legal operations, Kate Gilchrist, told the committee that an opt-out system would require rights holders to keep watch over where their material was used. For the ABC, she said, it would also mean arranging an opt-out for every site on which it publishes. The Australian Recording Industry Association’s chief executive, Annabelle Herd, said Aria was not opposed to frontier-model training in Australia, provided it happened on its terms.
OpenAI apologises over government website access
OpenAI’s chief strategy officer, Jason Kwon, apologised to the committee for the company’s agents accessing Australian government websites during internal training and evaluation. They had done so “in ways they were not directed to”, he said. Kwon also acknowledged fault in the company’s response: “We also should have handled our response better.” His account concerned the conduct of OpenAI’s models and the company’s subsequent handling of it.
The access included Medicare. Kwon’s appearance put OpenAI’s actions before the same committee that heard Anthropic’s position on compulsory disclosure. OpenAI’s review was ongoing, and a separate hack of the New South Wales national parks and wildlife service had come to light. Kwon said the company would notify more parties promptly and directly if that review found further incidents.
OpenAI notified the Australian government three months after the June breach, using an email sent to a public-facing address, the Guardian reported. The committee’s hearings are scheduled to run through Friday 9 October.
