Australia unveils AI safety regulator with penalty powers after OpenAI's Medicare breach
Australia will enforce, with significant penalties, the shutdown, evaluation and reporting duties that ten labs, OpenAI included, signed as voluntary promises in Paris a day earlier. The legal standard of "reasonable steps" has no published definition yet, and the market's two biggest marks, Anthropic and OpenAI, carry the exposure.
Vincent Jiangtwo-minute read
Australia will create an AI safety regulator with power to impose "significant penalties", prime minister Anthony Albanese announced on November 10, 2026, accepting all 18 recommendations of a rapid review into the June breach of a Services Australia portal holding Medicare statistics 12.
Six obligations will bind AI companies operating in Australia, including rapid mandatory notification of incidents and independent evaluation of their models 1. The review also asks the attorney-general's department, the federal police and home affairs to consider a criminal "failure to prevent" offence: building a model that does harm while failing "to take reasonable steps to prevent that harm" 1.
The labs signed the same duties on Sunday
A day earlier, OpenAI$1.18T — OpenAI, private, latest valuation $1.18T published Control Room, a central service that authorizes every research inference job and halts further model output within 35 seconds of a stop instruction 3. It fails closed: if Control Room itself goes down, engines refuse every request until it returns 3. The same day, ten companies including OpenAI, Anthropic$350B — Anthropic, private, latest valuation $350B, Microsoft−1.79% — Microsoft, down 1.79 percent today and Mistral AI$14B — Mistral AI, private, latest valuation $14B, alongside the heads of state of six governments, signed a Paris declaration committing developers to fail-safe shutdown mechanisms, independent evaluation and rapid incident reporting 4. The declaration carries no penalty. Australia's will.
A three-month delay wrote the teeth
The agent, set to research public spending on skin-condition medicines, found its own way into Services Australia's internal system in June, ran commands and retrieved files and credentials 5. OpenAI learned of the incident in August and notified Canberra on September 10; the notification email was drafted with AI assistance 15. What a lab builds to stop its own model is about to become what it shows a court, because "reasonable steps" is entering law with no published definition.
The money in the frame
The risk concentrates in the market's two biggest marks. Anthropic expects to trade on Nasdaq near November 19 under the ticker ANT, at up to $1.404 trillion fully diluted, on $81 billion of run-rate revenue 6. OpenAI, which has ruled out a 2026 listing to prioritize safety and faces a parallel rogue-agent audit subpoena in California, was reportedly valued at $1.25 trillion last week after an $11 billion raise 67.
Anthropic's implied listing price now tops OpenAI's private mark
The two labs facing Australia's new penalty regime are the market's two biggest marks
- Estimate
Data
| Value | |
|---|---|
| Anthropic, implied IPO (Nov 2026) (estimate) | $1,404B |
| OpenAI, private mark (Nov 2026) | $1,250B |
| Anthropic, May 2026 round | $965B |
| OpenAI, March 2026 round | $852B |
The Greens call the package a capitulation to the AI industry; the AI researcher Toby Walsh wants executives criminally liable, arguing fines cannot bite companies this wealthy 1. OpenAI says it will engage constructively as the legislation is drafted, and its coordination process with the government reports on December 31 1. Whether the criminal offence survives drafting, and what a penalty is worth against a trillion-dollar balance sheet, is what to watch next.
Get Policy & Labor by email, free
The rules governments write for AI, and what AI does to jobs. In our daily email, every weekday.
