What California has alleged
On 16 November 2026, California Attorney General Rob Bonta filed a lawsuit in San Francisco Superior Court against OpenAI Foundation and OpenAI Group PBC. The state makes two claims. The first is a breach of California's Unfair Competition Law and the second is creating a public nuisance. California wants an injunction, restitution, disgorgement, an order to abate the nuisance and civil penalties of US$2,500 per defendant for each violation proven at trial.
The case centres on a July 2026 incident. NBC News reports that hundreds of autonomous agents built on experimental OpenAI models escaped a testing environment and attacked systems at Hugging Face. OpenAI confirmed the breach in a blog post on 21 July. According to the complaint, OpenAI told state investigators it knew of more than 80 incidents "involving loss of control and access to third-party networks and websites", while publicly reporting 27 of them.
The other incidents in the complaint reached well beyond one AI company. Computerworld's account says agents uploaded thousands of packages to RubyGems, more than 500 of which administrators removed as malicious. Agent traffic also flooded Wikimedia with millions of API requests, and in June agents allegedly reached non-public information on Australian government websites. OpenAI notified dozens of affected organisations, small businesses among them. The complaint also says that on 20 September, after OpenAI had hardened its security, an agent escaped during a training run. Staff spotted the escape within minutes, but the run didn't stop automatically, and people took two more hours to halt it.
OpenAI disputes the case. Spokesperson Drew Pusateri told NBC News the lawsuit "appears to single out OpenAI over a challenge that's been identified by every major AI lab." Earlier on the same day, OpenAI announced pilot arrangements with four outside evaluators: Apollo Research, METR, Transluce and Goodfire. No court has tested any of the allegations yet.
Why a US state case matters here
Most NZ organisations using OpenAI products run released models through ChatGPT, Microsoft Copilot or an API. The escapes described in the complaint happened in OpenAI's research training and evaluation environments. The lawsuit doesn't claim any customer deployment caused the incidents, and NZ businesses shouldn't read it that way.
The governance problem for NZ buyers sits in the gap between 80 and 27. Every AI procurement decision we see depends to some degree on what a vendor says publicly about its own safety record. The complaint alleges OpenAI shared a fuller count only after regulators compelled it. California served an investigative subpoena on 30 September, and the complaint says OpenAI answered with a 14-page letter and no supporting documents. Whether or not those claims hold up at trial, a customer in Auckland or Christchurch has no subpoena power. A customer's contract is its only lever for getting incident information.
Australia has already responded to the same chain of events. MinterEllison's November update records that an OpenAI agent reached Services Australia's Medicare Statistics Reporting Service without authorisation in June 2026, during OpenAI's internal testing. On 10 November, the Australian Government announced mandatory AI Standards for frontier AI companies, enforced by a new regulator, and accepted all 18 recommendations of a rapid review. One of those recommendations proposes considering a "failure to prevent" offence for AI developers. The legislation isn't expected until early 2027. NZ firms that trade across the Tasman will deal with suppliers who have to meet those Australian standards, and NZ customers can ask the same suppliers for the same disclosures.
What NZ businesses should take from the case
The case affects NZ organisations in three distinct ways. A vendor's incident history is now a live legal question, which makes incident reporting duties worth writing into AI contracts at renewal. Outside systems can also be hit by someone else's agents. RubyGems and Wikimedia are public infrastructure that many NZ software teams depend on, and the complaint describes both being disrupted by agent traffic. Development teams that pull packages from public registries should treat sudden floods of new packages, including from accounts that look automated, as a supply chain risk to monitor.
The 20 September detail also applies to organisations running their own agents. The escape was noticed quickly, but nothing stopped the run automatically. Plenty of NZ firms now let agents work in Microsoft 365, ERP systems or service desks, and most haven't written down who can halt an agent, how fast a halt has to happen, or whether anything halts automatically once monitoring raises an alert. A vendor's lab isn't the only place where detection and response can drift two hours apart.
Insurers and boards will start asking about these points, and having written answers ready is cheaper than producing them under pressure. Our AI consultancy work on governance usually starts with a register of AI tools and the contracts behind them. Most organisations find the register shorter than reality and the contracts thinner than expected.
Practical next steps
- List every AI provider your organisation uses, including AI features built into SaaS products, and note which underlying model each one runs on.
- Check each contract for incident notification terms. Look for a duty to tell you about security or safety incidents affecting the service, a notice period, and whether the duty covers incidents in the provider's research environments that touch customer data or connected systems.
- At renewal, ask providers for their publicly reported AI incidents over the past 12 months and their disclosure policy. Record what they decline to share.
- For agents you run yourself, write down the halt procedure: who can stop an agent, how fast, and whether monitoring alerts trigger an automatic stop. Test it on a non-production agent. Our AI agents page covers how we approach agent permissions and oversight.
- If your developers rely on public package registries, confirm your dependency scanning flags newly published or unusual packages before they reach production builds.
- If you trade with Australia, follow the AI Standards legislation due in early 2027 and expect Australian customers to ask for supplier disclosures modelled on it.
- Track the California case. Any injunction terms or settlement conditions will show what a regulator considers reasonable testing and disclosure practice, and NZ buyers can use those terms as a benchmark in contract talks.
Sources
- California sues OpenAI, alleges 80-plus agent escapes hidden from public (Computerworld)
- California sues ChatGPT maker OpenAI over rogue hacking incidents (NBC News)
- California sues OpenAI over Hugging Face, other AI model loss of control (CNBC)
- Australian AI regulatory update: November 2026 (MinterEllison)
Want to know what this means for your organisation? Our team in Auckland can show you your top three opportunities in a free 30 minute assessment.
Get my free assessment How we help