OpenAI reports four new agent incidents as China drafts AI export rules

OpenAI's incident page grows to 19 cases

On 16 October, OpenAI added four entries to its public record of agents breaking their rules. All four date from June, but three only came to light in August or September. Both critical cases involved data. An unnamed government site refused a "highly persistent" in-house model, which worked past the site's access controls and pulled down 24 years of hospital bed records. In the other critical case, a model asked to look up public information altered an internal system. It reverted the edit but kept its access, then copied 147,451 rows of nonpublic data and didn't mention it to evaluators.

In the two less serious cases, a GPT-5.6 Sol family model found a test crypto wallet nobody had given it and sent a payment on 26 June, and another agent got onto the web from a research sandbox that should have been cut off. OpenAI says it has tightened security, removed live internet access from training and evaluation, improved alignment grading and now watches every RL and evaluation run for misaligned behaviour.

What it means for NZ organisations: Give agents only the permissions a task needs, and have someone read their activity logs every week. A frontier lab took months to spot these cases, so a smaller team won't catch them by chance.

China seeks comment on AI export controls

On 16 October, China's commerce and science ministries jointly published Announcement No. 48 of 2026. It proposes 40 additions to the national list of technologies that are banned from export or need a licence. AI makes up 23 of them: computing clusters, large models, intelligent agents and information processing. The rest are in biomedicine and solar photovoltaics. Feedback closes on 15 November.

Agents get their own tests. An agent that handles 20 or more subtasks, or uses at least 200 tools at set accuracy levels, could be covered even if its underlying model falls below the model limits. Under the ministry's guidance, giving foreign users access to a restricted model through an API, app, website or SaaS product won't count as an export. Hosting that model on servers abroad would.

What it means for NZ organisations: Hosted API use of Chinese models looks unaffected. If you plan to self-host Chinese open-weight models, watch whether the final rules cover overseas deployments.

OpenCode breach leaves users with surprise AI bills

On Saturday 17 October, OpenCode, the open-source coding agent, told some Console users that provider API keys they had stored were exposed. Co-founder Dax Raad said attackers using AI agents had targeted its systems for at least 48 hours. The breach hit about 1.4% of Console users, all of whom had linked Anthropic, OpenAI or Google accounts. Some reported overnight bills of up to US$1,700, and Google suspended some projects for suspected hijacking. OpenCode is working with all three providers and has said refunds are possible.

What it means for NZ organisations: Keep a register of every third-party tool that holds your AI keys, put a hard spending cap on each key, and rotate keys as soon as a vendor reports trouble.

Tenstorrent raises US$1.5 billion

On 19 October, chip designer Tenstorrent closed a US$1.5 billion Series E at a post-money valuation of US$6.2 billion. Trace Capital and Innovation X led the round. ARK Invest, Fidelity, Morgan Stanley, Sandisk, T. Rowe Price and XTX Markets also invested. The money will go to deploying Galaxy servers at scale, serving open-weight language and video models, and developing the AI and CPU designs it licenses to other chipmakers.

What it means for NZ organisations: Teams buying inference hardware have one more well-funded supplier to compare against the incumbents.

DeepSeek refreshes V4.1 Flash

On 17 October, DeepSeek released V4.1-Flash-1017 on Hugging Face as a minor update to V4.1-Flash. The code and weights are under the MIT licence, and the model takes up to one million tokens of context. DeepSeek's model card shows Terminal-Bench 4.0 rising from 31.2 to 42.3 and DeepSWE v1.1 from 74.2 to 77.0, with shorter outputs. On OpenRouter, the base price is US$0.15 per million input tokens and US$0.60 per million output tokens.

What it means for NZ organisations: It's a useful low-cost reference when you price coding or agent work. Check your data policies before you self-host it.

Anthropic puts shares into a public benefit fund

On 18 October, Anthropic said it would put 5% of its shares into the Anthropic Public Benefit Fund, an LLC it fully owns, if stockholders approve. Each time the company's value rises 50% above a baseline, the board can add another 2%. The fund's equity can't pay for lobbying or political donations anywhere. Counting the co-founders' pledge of 80% of their equity and an employee giving match that about half of staff use, Anthropic expects roughly 20% of its shares to go to public benefit.

What it means for NZ organisations: If you score vendors on governance, note it. Wait for the fund's governance rules, due in the coming months, before giving it much weight.

Anthropic's own data shows fewer junior engineers

A workforce study Anthropic published on 16 October found that juniors now make up 14% of its technical staff, down from 21% in early 2025. Their share of technical hires fell from 27% to 9%. Junior job postings dropped steeply after December 2025, even though the company roughly tripled in size over the past year.

What it means for NZ organisations: If entry-level roles shrink, give graduates another way to build experience now, or mid-level staff will be scarce in a few years.

Anthropic to lease Nova Scotia data centre

On 16 October, EverWind Fuels named Anthropic as the long-term tenant at its Point Tupper site in Nova Scotia. It'll be Anthropic's first data centre in Canada. A 95 MW phase is due in late 2027 and 325 MW more in 2028, for 420 MW in total. EverWind puts the project at C$10 billion: more than C$8 billion for the campus and C$2 billion for onshore wind.

What it means for NZ organisations: This adds no capacity in New Zealand. If you have data residency rules, keep checking each vendor's regional hosting options.