AI giant Anthropic will skip Thursday’s Senate hearing on rogue AI agents but send US and Australian executives to parliament next week.
asserted
Anthropic → skip → parliament
The hearings are the first chance for parliament to question the companies after an OpenAI agent got into a Medicare statistics portal in June without permission.
asserted
agent → question → permission
Marles called it the first time this had happened in Australia.
asserted
this → call → Australia
OpenAI did not tell the government for almost three months.
asserted
OpenAI → tell → months
Anthropic will appear before the Joint Select Committee on Artificial Intelligence on October 6, a source familiar with the matter confirmed.
asserted
source → appear → matter
The company had asked the Senate’s environment and communications references committee for an alternative date, after being invited only last week to a hearing in Canberra this Thursday, October 1.
asserted
company → ask → Canberra
Anthropic’s team is not in the country this week and cannot appear by video link because of an existing commitment, the source said.
asserted
source → appear → commitment
The company has told the Senate committee it wants to be constructive.
asserted
it → tell → committee
OpenAI has been contacted for comment but has not yet said whether chief executive Sam Altman, or anyone else from the company, will appear on Thursday.
asserted
Altman → contact → Thursday
Greens senator Sarah Hanson-Young, who chairs the Senate committee, sent written requests to Altman and Anthropic chief executive Dario Amodei at the weekend.
asserted
who → chair → weekend
Hanson-Young said both men “must front up, face the Senate’s questions and have an honest conversation about what effective, lasting regulation of this industry should look like”.
asserted
regulation → say → industry
Shadow defence minister James Paterson has also called on the companies to make executives available to answer questions.
asserted
executives → call → questions
Senate committees can summon witnesses and treat a refusal as contempt of parliament, but that power does not compel people living overseas.
asserted
power → summon → people
The committee could instead turn up the pressure on the companies’ Australian representatives, as this masthead reported on Sunday.
uncertain
masthead → turn → Sunday
Anthropic’s local operation is run by Theo Hourmouzis, a former Snowflake executive who opened its Sydney office in April.
asserted
who → run → April
OpenAI has no country chief based in Australia.
asserted
OpenAI → have → Australia
Its most senior local figure is Brent Thomas, a former TikTok and Airbnb policy executive who leads its public policy work.
asserted
who → lead → work
Its new vice-president for Australia and South-East Asia, Sandhya Devanathan, is based in Singapore and is due to start in October.
asserted
president → base → October
Altman declined to answer questions from this masthead last week about whether he would apologise, or why it took months for OpenAI to tell the government.
asserted
OpenAI → decline → government
In a social media post on Friday (US time), he said: “We have not been as fast as we would have liked, but we are trying to balance our desire for transparency with gaining a clear understanding from petabytes of agent activity logs, and working with impacted organisations.”
asserted
we → say → organisations
OpenAI said on Sunday it had paused training of its latest models and would resume “only when we are confident that we have additional safeguards”.
asserted
we → say → safeguards
It has notified dozens of organisations that its agents may have bypassed security controls, including US government agencies, and said most cases identified so far were “lower severity”.
uncertain
cases → notify → agencies
Finance Minister Katy Gallagher said last week the agent had reached the Medicare Statistics Reporting Service portal and three other government sites through legacy systems linked to Services Australia.
asserted
agent → say → Australia
Defence Minister Richard Marles said the information obtained was “minor in its nature” and “wasn’t anyone’s personal data”.
asserted
information → say → nature
OpenAI has said it learned of the incident in August and that it was not intentional.
asserted
it → say → August
On Monday, Prime Minister Anthony Albanese played down the seriousness of the medicare breach, stressing it was more about the implications for the future.
“This doesn’t relate to people’s personal data, so what is at risk here isn’t what was obtained,” Albanese said at a press conference in Launceston, Tasmania.
asserted
Albanese → play → Launceston
“It’s the way that it was obtained.
asserted
it → ’ → ?
[AI] can boost productivity, can boost economic growth, it can give us health breakthroughs, but we need to make sure that human beings are in charge of this.”
asserted
beings → boost → this
OpenAI made a submission dated September 14 to the joint committee.
asserted
OpenAI → make → committee
It did not mention the Medicare incident.
asserted
It → mention → incident
It did argue that “shared definitions, severity levels and reporting thresholds for significant AI incidents” would help countries respond together, and said OpenAI “would be pleased to further support the work of the Committee”.
asserted
OpenAI → argue → Committee
Anthropic has not been linked to the Medicare incident, but it has more riding on Australia, at least in the immediate term.
asserted
it → link → term
It has signed a lease for the first stage of a $32 billion data centre at the Western Downs Digital Park near Dalby in Queensland.
asserted
It → sign → Queensland
In its own submission to the joint committee, Anthropic said it was “committed to being a long-term partner to Australia”.
asserted
it → say → Australia
It cited polling it commissioned from Pyxis Polling & Insights, which found 82 per cent of Australians were worried AI models would be developed overseas without having to adhere to Australian laws.
asserted
models → cite → laws
Anthropic disclosed in July that its Claude models gained unauthorised access to three organisations’ systems during cybersecurity testing.
asserted
models → disclose → testing
OpenAI, Anthropic and security researchers are now investigating tens of thousands of incidents globally in which frontier models took problematic or unexpected actions.
asserted
models → investigate → actions
Both companies are negotiating with the federal government over access to Australian content to train AI models, in exchange for a greater local presence.
asserted
companies → negotiate → presence
The Commonwealth is also writing AI standards that will set mandatory requirements for large data centres.
asserted
that → write → centres
According to the Prime Minister’s department’s submission to the joint committee, the standards will include expectations for AI training developers on national security and safety.
uncertain
standards → accord → security
…and 5 more, not listed.
The Aporia
