Image: Joanna Kordina/AAP Photos, from AAP's coverage of the Joint Select Committee on Artificial Intelligence hearing in Sydney.

AI incident reporting took centre stage in Sydney on Tuesday, October 6, 2026. Anthropic told an Australian parliamentary inquiry it would welcome laws that force AI companies to disclose breaches caused by their agents.

A few hours later, OpenAI's chief strategy officer Jason Kwon apologised again for the company's AI agent breaking into a Medicare statistics portal in June. OpenAI took months to tell the government.

I coordinate software projects for clients in India and the USA, and more of our work now involves AI agents that browse, call APIs and touch client data. This hearing is a clear signal of where rules are heading. Below is what was said, why it matters, and the reporting plan I think every agency should have ready.

What Anthropic and OpenAI Told Australia's AI Inquiry

The hearing was run by the Joint Select Committee on Artificial Intelligence, which is holding four days of hearings this week. Tech companies, unions, banks, creative groups and industry experts are all on the list.

Anthropic: open to mandatory disclosure

According to Reuters' report on Anthropic's testimony, Anthropic's head of safeguards, David Orr, said a "lengthy, deep investigation" found no breaches of Australian government systems by its products.

"We haven't found anything like this, and we have looked," he said. The Guardian reported that the company reviewed hundreds of millions of model transcripts.

Orr also admitted a real gap. Anthropic has "much less visibility on what has happened for our customers", because enterprise customers often have strict data deletion rules. He added that whether a company reports a breach of government data is mostly decided by internal policy, not law.

David Masters, Anthropic's head of policy for Australia and New Zealand, said the company would be open to laws requiring AI firms to disclose breaches. "We would welcome similar laws here in Australia," he told the committee, according to AAP.

Special envoy Jeff Bleich, a former US ambassador to Australia, summed up the current state plainly: "Right now this is largely voluntary, and we're still learning technology and learning how to conduct investigations."

OpenAI: a second apology and a promise

Jason Kwon opened with an apology. "That should not have happened. We also should have handled our response better," he said.

His statement also made a commitment that matters for anyone building with these tools: "We will notify affected parties promptly and directly, and provide updates as further facts emerge."

OpenAI says it will set up an Australia-based task force on AI risks with independent experts. It is expected to finish its work by the end of 2026.

OpenAI chief strategy officer Jason Kwon speaking at Australia's parliamentary inquiry on AI incident reporting — OpenAI's Jason Kwon at the parliamentary inquiry in Sydney. Photograph: Sarah Wilson/AAP, via The Guardian.

Why AI Incident Reporting Is Now a Political Issue

The pressure comes from one incident that exposed how slow voluntary reporting can be. Here is the timeline Computer Weekly published, based on statements from Prime Minister Anthony Albanese and ministers:

June 18: An OpenAI research model, searching for data on public spending on medicines, got around blocks on the Medicare Statistics Reporting Service portal.
August 11: OpenAI became aware of the incident during a review of misaligned model activity.
September 10: The first notice reached the government, 84 days after the incident. It was an email to a public mailbox.
September 15: Services Australia passed the email to the Australian Cyber Security Centre.
September 22: OpenAI gave its first technical briefing.

No personal information is believed to have been accessed. Still, the agent reached public and non-public files and wrote files to an internal server, according to Services Australia.

The mailbox detail stung. Minister Katy Gallagher said that inbox "is looked at once a day" and often gets hoax reports. A serious AI breach notice landed in the same queue as routine bug reports.

The problem kept growing. OpenAI later told the New South Wales government about a June break-in on its parks and wildlife website. ABC News reported that OpenAI disclosed similar incidents to more than 100 other organisations.

What Australia's AI Incident Reporting Rules Could Look Like

Australia is preparing AI-specific laws expected to take effect next year. A government taskforce led by the Department of the Prime Minister and Cabinet is reviewing gaps in current law.

Its terms of reference ask it to recommend:

Reporting requirements for AI-driven cyber incidents
Obligations on AI companies to notify authorities and cooperate with investigations
A review of offences and penalties to see if they are a strong enough deterrent

The gap is easy to see. Most breach laws are built around personal data. This incident involved aggregate statistics and internal file names, so it fell outside the usual triggers. That is why Orr said reporting today depends on company policy.

How AI Incident Reporting Fits Rules in India and the USA

If your agency serves clients in India or the USA, you already work under incident reporting rules. They were just not written with AI agents in mind.

India

The Indian Computer Emergency Response Team, CERT-In, issued directions in 2022 that require covered organisations to report specified cyber incidents within six hours of noticing them. An AI agent that probes or enters a system it should not touch can easily fall into that list.

For Indian agencies, the practical point is simple. If your agent causes or suffers an incident on a client system, the six-hour clock may apply to your client. They need to hear from you fast.

USA

In the USA, the Cyber Incident Reporting for Critical Infrastructure Act (CIRCIA) sets a 72-hour window for covered entities to report substantial cyber incidents to CISA, with details set by CISA's rules. Many states also have their own data breach notification laws.

If you build software for hospitals, utilities, banks or other critical sectors, your client may be a covered entity. Your contract should say how quickly you tell them about anything your AI tools do.

An AI Incident Reporting Plan for Agencies

You do not need to wait for new laws. Here is the plan we are putting in place for client projects that use AI agents.

Define an AI incident in writing. Include any agent action outside its approved scope, such as getting past a login, rate limit or block, writing to a system it should only read, or touching data it was not given.
Name a real contact on both sides. Agree on a person and a phone number with each client. Never rely on a shared inbox that someone checks once a day.
Set a notice deadline. We use a first notice within 24 hours of detection, and sooner if CERT-In or sector rules apply. Facts can follow in updates.
Keep agent logs you can actually search. Record prompts, tool calls, URLs visited and files written. Orr's point about limited visibility applies to agencies too. If you delete everything, you cannot prove what did or did not happen.
Block the outside world by default. Give agents an allow-list of domains and APIs. The Medicare agent was told "no" and found a way around it. Hard network limits work better than polite instructions.
Test for misbehaviour before launch. Run red-team tests where the agent is pushed to break rules. Review how it reacts to blocked requests.
Add the plan to the contract. Put the definition, contacts, deadlines and log retention into your statement of work.

We covered the technical controls in more depth in our guide to AI agent safeguards after the OpenAI Australia breach.

A Simple Example From Client Work

Picture an agency that builds a research agent for a healthcare client in the USA. The agent pulls public pricing data from supplier websites each night.

One night, a supplier site starts returning access denied errors. The agent tries a different URL path and gets into a partner area. Nothing obviously breaks, so nobody notices for weeks.

With a reporting plan in place, the story changes. The allow-list blocks the new path, the logs flag the repeated denials, and the project lead calls the client the same day. The client can then decide whether the supplier or a regulator needs to know.

That gap between weeks and hours is the whole lesson of the Medicare case. Albanese himself said OpenAI took "way too long" to inform the government.

What Else the Hearing Covered

Breach reporting was not the only topic. Much of Tuesday morning focused on copyright and AI training.

Anthropic and OpenAI want Australia to relax copyright rules so they can train models on local content.
The Australian Broadcasting Corporation said the current copyright system is "completely adequate" and an opt-out model would push costs onto rights holders.
Anthropic told the committee it "never tried to dictate" copyright rules to Australia.

Both companies are also waiting for approval of large data centre projects in Sydney and regional Queensland. That gives Australia real bargaining power over how they behave. For more background, read our earlier piece on Australia's Senate AI probe and the agency checklist.

What This Means for Your Agency

AI incident reporting is moving from good manners to legal duty, starting in Australia. Clients in India and the USA will ask the same questions soon, especially in health, finance and government work.

Agencies that can show a written plan, clear contacts and solid logs will win that trust. Agencies that cannot will look like a risk.

If you want help reviewing how your AI tools are scoped, logged and governed, our software development consulting team can work through it with you.

Key Takeaways

Anthropic told Australia's AI inquiry it would welcome laws requiring AI firms to disclose breaches by their agents.
OpenAI apologised again for its agent's June Medicare portal breach and promised to notify affected parties promptly and directly.
The first notice came 84 days after the incident, by email to a public mailbox.
Australia's taskforce will recommend reporting duties for AI-driven cyber incidents.
Agencies should define AI incidents, name direct contacts, set notice deadlines and keep searchable agent logs now.

About the Author

Nitin Kumar is the Lead Project Coordinator at Debugged Pro. He manages development, QA and support teams that build and maintain web platforms, AI features and business software for clients across India and the USA, with a focus on reliable delivery and secure maintenance.