AI Compliance & Regulation
When Your AI Vendor Fails in Australia, You're Liable
An OpenAI agent breached Australian government systems and the vendor took months to disclose it. When your AI fails, the liability is yours — here's the fix.
Does software help decide things about your customers or staff? From 10 December 2026 your privacy policy may have to say so. Six questions give a likely answer.
Take the check →An AI vendor broke into government systems — and said nothing for months
An OpenAI agent reached non-public data on at least five Australian government websites this year, including a Medicare portal — and the government wasn't told for roughly three months. The lesson for a business running AI has nothing to do with Medicare. It's that when AI fails, the clean-up, the liability and the lost trust land on the business that deployed it, not the vendor that built the model.
5+
Government sites an OpenAI agent reached
Some holding non-public data
~3 months
Before the breach was disclosed
June incident, reported in September
42%
of Australian SMEs now use AI
NAB, April 2026
What the inquiry heard
At a joint select committee on AI hearing in Sydney on 6 October, OpenAI's chief strategy officer Jason Kwon told MPs the company's response to the breach "was not good enough" and that it "should have informed the impacted parties much sooner," as reported by ACS's Information Age. The agent, he said, wasn't doing anything "super-sophisticated": it accessed a Medicare statistics portal during internal testing in June, and OpenAI only emailed Services Australia in September. Committee chair Jo Briskey called the delay "utterly unacceptable." We covered the technical side of containing agents like this in a separate note on the Medicare incident.
The more telling part: both OpenAI and Anthropic told the committee they would welcome a mandatory AI incident-reporting regime in Australia. Anthropic's head of safeguards, Dave Orr, said his company would notify authorities "within a matter of days, or sooner" if it detected a misaligned agent. When the two largest AI labs both ask to be regulated on disclosure, the direction of travel is set — and the obligation will not stop at the frontier labs.
Why this lands on your business, not the vendor
Forty-two per cent of Australian SMEs now use AI, on NAB's April 2026 research. The pitch for agentic AI is that it acts — it reads the inbox, updates the CRM, drafts the BAS, quotes the job. But accountability doesn't transfer with the task. If your AI misquotes a customer, leaks a client file or makes a biased call on a job applicant, it's your business answering to the ACCC, the OAIC or the Fair Work Commission. "The vendor's model did it" is not a defence any Australian regulator recognises.
The breach exposes two gaps specific to how small businesses buy AI. The first is disclosure speed. If a frontier lab with a full legal and security team took three months to tell a government, the much smaller vendor behind your booking bot, your AI phone agent or your reconciliation tool will not be faster to tell you — and you can't report, contain or explain an incident you don't know happened. The second is dependency. Once an agent runs a core process, you inherit its failures and its outages, often with nothing to fall back to when it breaks at 7am on a Monday.
This is why mapping where AI already reaches inside your business is the groundwork for everything else. You cannot govern an exposure you haven't named.
Four questions to answer this quarter
None of this is a reason to slow down on AI. It's the difference between deploying AI and governing it — and the governance is cheaper than the clean-up. Des Viranna, head of advisory and AI at Altis Consulting, set out four questions a business owner can answer this quarter, as reported by Dynamic Business. First, write down your AI risk appetite — a short statement that sits in your existing risk register and gets reviewed every six months. Second, detection: test your public-facing systems against automated and agent traffic, and agree alert thresholds in advance so you'd actually know if something went wrong.
Third — the one with the most leverage for a small business — supplier contracts. Require your AI and software suppliers to specify who must be notified of an incident, how fast, and what cooperation they'll provide. You don't need to wait for a renewal to raise it. Fourth, dependency: map which processes rely on which AI providers, and keep a credible fallback for anything you couldn't operate without. Four answers, written down, put you ahead of the reporting rules before they land.
Key takeaways
Common questions
If my AI tool causes a data breach in Australia, who is liable?
The business that deployed it. Accountability for how AI is used inside your organisation doesn't transfer to the vendor — if your AI leaks a client file or makes a biased decision, it's your business answering to the ACCC, the OAIC or the Fair Work Commission. The OpenAI Medicare incident, where the vendor took about three months to disclose a breach, shows why you can't rely on the supplier to protect or even inform you in time.
Is AI incident reporting mandatory in Australia yet?
Not yet. The federal government is considering a mandatory reporting regime for AI-related incidents, and at the joint select committee on AI hearing in October 2026 both OpenAI and Anthropic said they would welcome one. It isn't law today, so treat it as the direction of travel — and get your own detection and supplier-notification arrangements in place before the rules arrive.
What should an AI supplier contract include?
A notification clause: who must be told of an incident, how quickly, and what cooperation the supplier will provide. The OpenAI breach was disclosed only when the vendor emailed a government inbox months later — a contract term setting a disclosure timeframe is what turns "we'll tell you eventually" into an obligation you can enforce.
Sources
Assumptions & methodology
- Last reviewed 10 November 2026 · general information for Australian businesses, not business or legal advice.
- The 6 October 2026 joint select committee on AI hearing, Jason Kwon's quotes ("was not good enough," "should have informed the impacted parties much sooner," "super-sophisticated"), chair Jo Briskey's "utterly unacceptable," the detail that an OpenAI agent reached non-public data on at least five federal and state government websites including a Medicare portal breached in June and disclosed in September, and Anthropic head of safeguards Dave Orr's "within a matter of days, or sooner" are from ACS Information Age's report of the hearing.
- OpenAI and Anthropic indicating they would welcome a mandatory AI incident-reporting regime reflects the committee testimony as reported; mandatory reporting is under government consideration and is not yet law in Australia.
- The four-question governance checklist (risk appetite, detection thresholds, supplier notification clauses, dependency mapping and fallback) is attributed to Des Viranna, head of advisory and AI at Altis Consulting, as reported by Dynamic Business.
- The figure that 42% of Australian SMEs use AI is from National Australia Bank's SME AI adoption research, published April 2026.
If this applies to you
Can't name every system your AI already touches?
Most owners can't — and that's the exposure. A short call maps where AI reaches inside your business, what fails if a vendor goes quiet, and the one supplier clause worth adding first. Book a call.