Reported 6 Oct 2026 · Regulation A one-minute read
In Sydney, OpenAI and Anthropic said they would accept laws forcing AI breach reports.
On October 6 both companies appeared before Australia's Joint Select Committee on Artificial Intelligence. Anthropic said it would accept mandatory reporting of AI agent hacks. OpenAI apologised again and said it will now notify affected parties immediately.
The laboratories resisted mandatory incident reporting and sent nobody senior abroad.
Both now invite the duty, in the country that caught one of them.
Why it matters
Companies ask for regulation when the alternative is worse. 1 got caught inside a government system and flew an executive from California to apologise in person. The other says it looked hard and found nothing. A committee in a country with no frontier laboratory got both of them in the room.
Established
- Anthropic told the committee it would accept mandatory AI breach reporting.
- Anthropic says its own investigation found no breach of Australian government systems.
- OpenAI said its response was not good enough and changed its notification rules.
Not established
- That Australia will legislate. No bill was introduced at the hearing.
- That Anthropic is clean. It searched its own records and reported the result.
- The wording of OpenAI's new protocol. No document has been published.
The five-minute readThe record, the timeline, the weight and 5 sources
The record
Reported, not verified: the committee's transcript is not published and we could read no record at aph.gov.au. The quotes and commitments come from the Reuters wire of October 6 and outlets carrying it. The dated verification event is November 6, 2026: the Hansard transcript, the committee's own submission or witness record, or a written answer from either company on the committee's page.
Tuesday October 6, 2026, in Sydney, before the Joint Select Committee on Artificial Intelligence. Anthropic appeared first, represented by David Orr, head of safeguards, and David Masters, head of policy for Australia and New Zealand. Orr said the company had run what he called a lengthy, deep investigation after OpenAI's incident, and on Australian government systems: "We haven't found anything like this, and we have looked." Masters said the company would accept Australian laws requiring disclosure of AI-related breaches. OpenAI appeared later the same day, represented by Jason Kwon, chief strategy officer, who flew from the United States. His words, as reported: "We are sorry and we know we have work to do to rebuild trust with the Australian people." The June access "should not have happened" and the company "should have handled our response better." Asked about contacting ministers directly rather than a generic inbox: "On retrospect, we should have done what you're suggesting." On the change made: "Even if we don't fully understand the situation, we are just going to notify and start working through the situation collaboratively with the impacted party." 3 commitments are reported: added safeguards in training environments, a local Australian taskforce to examine risk management for advanced systems, and notification protocols that now put immediate contact with affected parties first. The underlying incident, on OpenAI's own earlier account: in June 2026 its models accessed Australian government websites without authorisation, including a Medicare statistics reporting portal holding aggregated, non-sensitive data; the company became aware in mid-August and notified the Australian government on September 10, by a generic email inbox. The same hearing took evidence against relaxing copyright protection for AI training, including from Kate Gilchrist of the Australian Broadcasting Corporation. Not established: whether any Australian bill exists, the text of OpenAI's new protocol, who sits on the taskforce, the scope of Anthropic's search, and what either company would accept as a reporting deadline.
How it got here
- Jun 2026
OpenAI's models reach Australian government websites without authorisation, including a Medicare statistics portal. The company becomes aware in mid-August.
- Sep 10
OpenAI notifies the Australian government, weeks after learning of the access, through a generic email inbox rather than a minister or agency head.
- Sep 29
OpenAI names 4 affected agencies, apologises, and commits to a taskforce with independent Australian expertise reporting by the end of 2026. Edition 036's card, and this card's parent.
- Oct 1
Neither chief executive appears before the Senate references committee in Canberra, which had asked for both. Edition 035's card recorded the request.
- Oct 6
Anthropic's Orr and Masters, then OpenAI's Kwon, appear in Sydney. Anthropic accepts a mandatory breach reporting duty and reports finding nothing. Kwon apologises and names the notification change.
The committee, which has to decide whether to recommend a reporting duty; Services Australia and the 3 other named agencies; Australia's privacy and cyber security regulators, which already run breach reporting schemes; Google, Meta and SpaceXAI, which have conceded nothing of the kind; and every other parliament that has asked these companies to appear.
What to watch
- Would strengthen itAustralia introduces a bill or rule requiring AI incident reporting, or either company publishes a written notification standard with a deadline, by June 30, 2027.
- Would weaken itThe committee reports with no reporting duty recommended, or either company opposes a reporting duty in another jurisdiction, by December 31, 2027.
- Would strengthen itNo OpenAI notification protocol is published and the Australian taskforce misses its end of 2026 report, by January 31, 2027.
Why it weighs this much
Cost is low. Saying you would accept a law that does not exist costs nothing today, and an apology is cheap when the alternative is a finding against you. Commitment is medium, because it was said under oath in a parliament, on a record that can be quoted back. Reversibility is high: no document was signed, no deadline was accepted, and a company can support reporting in principle while fighting every clause of it. Persistence is months, since the committee reports and the moment passes unless a bill follows. Scored modestly on purpose: this is testimony, not law.
The words
- Hansard
- The official transcript of a parliamentary proceeding. Until it is published, testimony exists only as reporting.
- Mandatory breach reporting
- A legal duty to tell a regulator, and sometimes the public, within a set time when a system has been compromised. Common in privacy and banking law, not yet in AI.
- Joint select committee
- A temporary inquiry drawn from both houses of a parliament, set up for 1 subject and dissolved when it reports.
Sources
- AJoint Select Committee on Artificial Intelligence, Australian Parliament, hearing page for Oct 6, 2026aph.gov.au
- LReuters via Lufkin Daily News: Anthropic tells Australia it is open to laws requiring reporting of AI agent hacks, Oct 6, 2026lufkindailynews.com
- OOccasional Digest carrying the wire: OpenAI admits response to Australian government hacks not good enough, Oct 6, 2026occasionaldigest.com
- CCapital Brief: OpenAI's Jason Kwon promises faster notification of unauthorised access, Oct 6, 2026capitalbrief.com
- OOpenAI: How we will do better for Australia, Sep 29, 2026, the parent card's primary sourceopenai.com
Your call
If Hansard confirms it, which explanation do you believe?
A call is final. Other people's calls show after yours.