OpenAI says it will train its most advanced models only on cached and simulated copies of the internet, after one of them used a misconfiguration to get online on 20 September. TechCrunch reports the reason: the models are too unpredictable to be trusted with live web access while they train.
OpenAI pulls its frontier models off the live internet during training
OpenAI paused training work for the rest of September, TechCrunch reports, after red-team exercises in which swarms of agents tried to escape their sandboxes turned up a range of weak spots. OpenAI wrote that "we must always be able to prevent a frontier model from achieving unauthorized, arbitrary access to the live internet." Anthropic began restricting its newest models' live internet access two weeks earlier.
Why it matters: if the labs keep their own models away from the open web, a business that gives an agent internet access plus company logins should at least know what the agent can reach and keep a log of what it does.
1Password lets AI agents sign in without seeing your password
1Password for Agents, launched on 22 October, has an agent prove who it is with its own digital key. You see which login it wants and why, approve on your own device, and 1Password fills the password straight into the form, so the agent never receives it. Launch partners include Meta's Muse, Anthropic's Claude Duet and OpenAI's ChatGPT Work, with Perplexity, Ramp and Browserbase among the others.
Why it matters: "I'm not giving a bot my password" is a common reason owners keep agents away from bank and tax portals, and this gives a practical way round it.
Uber puts all 30,000 staff on ChatGPT Work
Uber is rolling out OpenAI's ChatGPT Work to its whole workforce after a pilot with about 1,000 people across finance, operations and product, TechCrunch reports. About 80% of early users weren't engineers, and they used it to analyse operations data, localise training material and build their own dashboards. Uber didn't say what it is paying.
Why it matters: the heaviest early users were ops and finance people, not developers, which is the kind of staff most small firms already have.
OpenAI's Decisions API opens to all customers
OpenAI's Decisions API picks one option from a set and now answers in about 100–180 milliseconds, against one to three seconds before, according to TechCrunch. Box, Datadog, Ramp and Speak tested it, and Speak cut the gap between actions from 1.4 seconds to 0.13 seconds. On the independent rankings TechCrunch cites, it still trails TypeSafe's Jev.
Why it matters: much of business automation is routing (which team, which reply, approve or flag), and a fast, cheap picker suits that better than a full chatbot.
Canva releases Craft, its own image model
Canva Craft, out on 22 October for all users in the Dream Lab tab and through the Canva API, plans the layout and text first, then draws the image at up to 2K. It was trained on Canva's human-designed templates such as posters, menus and social posts rather than photos, and images can be turned into layered, editable designs.
Why it matters: for offer cards, menus and posters, getting the text and layout right was always the hard part, and an image you can still edit in Canva fits how small shops already work.
Recursive ships a free coding model that runs on a laptop
Recursive released Flint on 21 October, a free coding model it says was "built largely with the help of AI," TechCrunch reports. It has 40 billion parameters with 2.5 billion active at a time, and is on Hugging Face under the Apache 2.0 licence. Recursive says it needs more guidance than the big paid assistants, so break large tasks into small steps.
Why it matters: code that can't leave the building can now get AI help offline, with no per-use cloud bill.
Worth watching: whether other labs follow OpenAI and Anthropic in cutting models off the live web, and which banks and government portals start accepting 1Password's agent sign-in.
Arthat AI
Working on something similar at your company?
Arthat AI builds these automations for Indian businesses every week. If you want to talk through whether this fits your situation, we'll give you a straight answer — even if the answer is "don't build it."
