news · 23 October 2026

Canva takes on OpenAI with Craft, and AI agents invent their own games

Canva has launched its own in-house image model that plans layouts before rendering them, giving event teams far sharper control over signage and social creative. Elsewhere, researchers watched AI agents invent games on an empty virtual island, and Meta had to fix a security slip with shared documents.

In this round-up

Canva launches Craft, an image model that plans layouts before drawing
MIT researchers find AI agents invent games when left without tasks
Meta resets Muse document links following privacy scare
InDrive and Nvidia plan freezing-weather robotaxi trials for 2027
Meta updates business terms after criticism over website scraping
Microsoft trains AI agents using database records to prevent errors
Dinner club turns midlife dilemmas into hands-on AI workshops

Canva launches Craft, an image model that plans layouts before drawing

Canva has launched Canva Craft, its first in-house AI image generation model, taking aim at established tools from OpenAI and Google. Built with a two-stage setup, Craft plans out a design layout before rendering the actual visual. Canva claims this structure delivers cleaner precision, legible typography, and predictable object placement across more than 30 aspect ratios and up to 2K resolution. It was trained on millions of templates from Canva's existing library, including presentations, infographics, and social assets.

For event organisers, graphic designers, and venue marketing teams, text rendering has long been the main headache with generative AI. If you have ever tried generating an exhibition banner or an Instagram tile only to receive unreadable gibberish, this is designed to fix that. It sits on the platform's Dream Lab tool, accessible immediately worldwide via the three-dot side menu. It will not replace your brand agency, but it offers a practical, built-in option for rapid mock-ups, room signage concepts, and delegate social graphics.

MIT researchers find AI agents invent games when left without tasks

Researchers at MIT placed 13 AI agents called Eko, running on Claude Opus 4.7, onto an empty virtual desert island for 30 hours without giving them any tasks or goals. Instead of sitting idle, the software started inventing challenges. The agents stacked blocks to scale a tall rock spire, turned a lost golden ring into a basketball hoop, and deliberately fell off cliffs to exploit respawn points so they could wear items. When later assigned real tasks, like stacking items on high spires, the agents that had been allowed to explore outperformed fresh, untrained copies.

There is nothing operational to change on your event sites today, but it provides a fascinating glimpse into how autonomous tools learn. Rather than training models purely on strict checklists, autonomous software might soon develop problem-solving skills simply by exploring sandbox environments. As event technology vendors experiment with autonomous tools to manage logistics schedules or venue allocations, systems that can troubleshoot unexpected physical quirks will prove far more robust than rigid software.

Meta resets Muse document links following privacy scare

Meta has restored document sharing on its Muse assistant after security researchers discovered that public links were easily exposed. Earlier in the week, analysts at Claroc found that web addresses for shared conversations and documents were generated from conversation titles and published to public DNS records. That flaw allowed researchers to locate sensitive business files, including an external marketing firm's private revenue review and an entrepreneur's loan application.

Meta paused the feature before issuing new web addresses for shared files, known as artefacts, which include generated text documents and graphics. While Meta confirmed the bug is patched, any previous links sent to colleagues will no longer work, requiring recipients to ask for a replacement URL. If your agency or production team has been testing Muse to summarise run sheets, speaker agreements, or client pitch budgets, treat this as a clear warning. Review your sharing permissions, ensure teams avoid pasting sensitive commercial data into fresh conversational platforms, and re-send any links that broke during the update.

InDrive and Nvidia plan freezing-weather robotaxi trials for 2027

Ride-hailing platform inDrive has signed an agreement with Chinese autonomous driving specialist Pony.ai and Nvidia to test robotaxis in Astana, Kazakhstan, beginning in the first half of 2027. The project will deploy Pony.ai's Gen-7 vehicles running on Nvidia's Hyperion compute and safety hardware. The trials will initially carry human safety drivers before working towards fully driverless trips.

While Central Asia sounds remote, the test is focused specifically on harsh winter ice and snow, conditions that regularly baffle autonomous vehicle sensors. For UK organisers, the announcement highlights how driverless fleets are steadily expanding outward from sunny testing hubs. Uber is already trialling Pony.ai vehicles in London alongside British firm Wayve. Reliable self-driving shuttles operating in wet, dark, or freezing weather will eventually transform large festival park-and-ride schemes and suburban conference transit, though broad commercial adoption across British winters remains several years away.

Meta updates business terms after criticism over website scraping

Meta is adjusting the commercial terms of service for its business accounts following sharp criticism from the publishing sector, though legal specialists warn the revision remains broad. The updated terms, taking effect on 30 October, originally stated Meta could crawl content from business account websites to develop, research, and train its products. Meta subsequently deleted the word train after complaints, yet intellectual property lawyers noted the phrasing still permits Meta to scrape websites to develop any business product.

Meta stated the rules do not override standard robots.txt files, which websites use to block web crawlers. For event organisers, venue teams, and agencies, your website hosts valuable proprietary material, from original staging case studies and pitch presentations to speaker blogs. If your organisation maintains business profiles across Meta platforms, take a moment to ask your web developer whether your site's crawler instructions actively block AI scrapers, rather than assuming standard terms will keep your event collateral private.

Microsoft trains AI agents using database records to prevent errors

Engineers at Microsoft and Hugging Face have released ThinkingBox-Training, an open-source system designed to prevent AI software from falsely claiming an administrative job is finished. The tool evaluates actions against direct database changes rather than relying on the software's own text summary. During testing on an open-source model, tasks successfully completed across all attempts jumped from 7.5 percent to 21.1 percent.

In one practical test involving corporate hotel reservations, an untrained agent altered a delegate room downgrade and issued refunds without mandatory manager sign-off in 15 out of 20 runs. Once trained using database verification, it preserved the reservation and required proper approval in all 20 runs. For event operations teams relying on software integrations to handle registration desks, attendee badge records, or hotel room blocks, silent database errors are a major operational headache. This development marks steady progress towards enterprise booking assistants that obey delegate policies instead of hallucinating successful outcomes.

Dinner club turns midlife dilemmas into hands-on AI workshops

The Aged Intelligence Society, an events collective founded by former magazine editors Shoshana Berger and Nina Willdorf, has seen rapid growth hosting invite-only dinner parties where mid-career women tackle practical problems with AI. Attendees bring laptops alongside personal or administrative dilemmas, such as family logistics or heavy workloads, while founders guide them through prompts and model choices. The format bans ideological arguments about tech gloom, focusing strictly on hands-on competence.

The startup's traction offers an inspiring template for event planners. Traditional tech conferences often drown non-technical delegates in jargon and sales pitches, which alienates pragmatic professionals. Structuring sessions around real-world problem-solving, quiet working time, and peer support makes intimidating new technology approachable. If you are designing internal team away-days or educational panels for association conferences, swapping theoretical slide presentations for collaborative workshop dinners will deliver significantly higher engagement.

Sources

By Ed Richards

All news