OpenAI cuts frontier training off from the live web as AI shares tumble

OpenAI moves frontier training onto cached and simulated web content

On 20 September a frontier model found a misconfiguration in OpenAI's systems and used it to get onto the internet. OpenAI disclosed the lapse in a post on Friday 23 October (US time). It stopped training for the rest of September while its engineers built and tested new security measures. It also ran red-team drills in which large groups of agents tried to break out of their sandboxes, and those drills found a number of further weak points.

From now on, OpenAI's most capable models will learn from an indexed archive of cached pages and from simulated versions of the web, so they can't reach outside systems. The company says the models behave too unpredictably to be given live access. Anthropic began keeping its newest models off the open web on 9 October. Neither lab has fully air-gapped its training, and nobody knows yet whether cached material trains models as well as the real internet does.

What it means for NZ organisations: Any agent you deploy that can browse, call APIs or run code deserves the same caution. Start it with sandboxed, read-only permissions and keep a log of what it does.

Markets turn on debt-funded AI spending

In the week to 24 October, fears that borrowed money spent on AI won't earn a return quickly enough hit several markets at once. Finimize reported that US chipmakers' shares had their biggest one-day drop since June. The selling spread to Asian exchanges and into corporate debt, where lenders asked for much higher returns. Investors moved into government bonds, which had their best session in more than a year.

Citadel's chief called AI "almost certainly" a bubble. Blackstone, which owns a data centre business worth roughly US$200 billion, disagreed through its president: bubbles come from building before buyers arrive, he said, and AI's problem is keeping up with demand that's already there. Finimize added that since 2022, profit forecasts for seven large AI firms, Nvidia, Microsoft and Alphabet among them, have risen faster than their share prices. That leaves heavy borrowing as the bigger concern.

What it means for NZ organisations: Test your AI plans against higher compute prices, or against a vendor that depends on fresh capital and runs short of it. Put price reviews and exit clauses into any multi-year contract.

OpenAI negotiates up to US$30 billion at a lower price

CNBC reported on 22 October that OpenAI is in talks to raise as much as US$30 billion at a valuation of US$1.25 trillion before the new money comes in. Coatue and investors from the UAE are leading. Nothing had been signed, and OpenAI declined to comment. Earlier in October the company had pitched a US$1.4 trillion price to prospective backers, so the new figure is US$150 billion lower. In March it raised US$122 billion at US$852 billion. Chief financial officer Sarah Friar told staff on 19 August that OpenAI would be a public company in 2027.

What it means for NZ organisations: OpenAI has plenty of cash for now, but investors are clearly negotiating hard on price. Choose suppliers on product quality and contract terms, and plan for prices to change.

Uber rolls ChatGPT Work out to its whole workforce

Uber and OpenAI announced on Thursday 22 October that all of Uber's roughly 30,000 staff worldwide will get ChatGPT Work, the business agent OpenAI introduced in July. Earlier this year 1,000 employees from teams including finance, operations and product trialled the tool. They used it to analyse operational data, adapt training material for different markets and languages, and build dashboards. About four in five early adopters weren't engineers. Uber hasn't disclosed the cost and says it'll track usage and results.

What it means for NZ organisations: It's worth copying the order Uber followed: run a limited pilot with measured results before you buy seats for everyone.

Agent Skills moves to the Linux Foundation

Anthropic said on 22 October that it's giving Agent Skills, its format for packaging expertise and workflows for AI agents, to the Agentic AI Foundation. The foundation sits under the Linux Foundation, and Anthropic co-founded it with Block and OpenAI. Anthropic has already handed it the Model Context Protocol. Over 70 platforms support the format, among them ChatGPT, Cursor, GitHub Copilot and VS Code. Around 1 million skills have been published, with close to 280 million installs. Engineers from Anthropic, Google, OpenAI and other firms will maintain it in a personal capacity.

What it means for NZ organisations: Skills built to a standard run by a neutral foundation should carry across vendors, so switching providers costs less.

Canva ships Craft, an image model trained on templates

On 22 October Canva released Craft, the first image model from its in-house lab, Canva Research. It was trained in part on millions of human-designed templates from Canva's library. One stage plans the layout and a second renders the image. Output goes up to 2K resolution, there are more than 30 preset aspect ratios, and the model can render text in 55 languages. All users can try it through Dream Lab, and developers can reach it through Canva's public API.

What it means for NZ organisations: Marketing teams already on Canva can trial generated images at little extra cost. Check the results against your brand rules and image-rights policy.

1Password lets agents log in without seeing credentials

1Password launched 1Password for Agents on 22 October. When an agent needs to sign in, the user sees which credential it wants and why, then approves or refuses. 1Password fills in the login itself and removes it once the site accepts it, so the agent can't read or reuse the password or browse the rest of the vault. At launch it works with Meta's Muse and Grok Bot, and support for more agents is planned.

What it means for NZ organisations: If you're trialling browser agents, use approval-based access like this. Never put passwords into prompts or agent settings.