🏦 Major US banks restrict AI agents in customer accounts

🎨 Canva launches Craft, its first AI image generator

🩺 Microsoft AI beats doctors in diagnostic tests

🚗 Uber just gave 30,000 workers an AI coworker

🤖 South Korea bets on robots to fill its worker gap

+ 📊 Daily poll and results

+ 📈 Trending tools and resources

Major U.S. banks caught third-party AI agents logging into customer accounts with customers’ own credentials, Axios reported. Morgan Stanley is blocking them outright, while JPMorgan, Bank of America, and Citi are restricting access.

Most agents had user permission, but banks spotted some calling support lines or acting beyond instructions. Executives called it “horrifying,” and one admitted banks struggle to tell authorized agents from rogue ones.

Most banks are landing on read-only access, the same deal budgeting apps like Quicken already get. So your AI agent can check your balance, but paying the bills is still your job.

Canva launched Craft, its first image generation model, letting users create images up to 2K resolution in 30 aspect ratios through Dream Lab, with access through its API also coming.

Unlike photo-focused models, Craft learns from Canva’s design templates and plans layouts before generating images. Canva says this approach improves text accuracy, enables precise edits, and supports converting images into editable designs.

Canva has relied on outside AI image models for years. Building its own could lower API costs and give it more control over speed and quality, especially as AI expenses pressure growth.

Microsoft’s MAI-DxO AI system outperformed doctors on complex medical cases and simulated emergency room visits, while ordering fewer and cheaper tests, according to the company’s newly published results.

The system uses multiple AI agents to debate diagnoses and select tests. It achieved 80% accuracy on typical ER cases, compared with 47% for physicians using their usual diagnostic approaches.

Strong test scores don't guarantee better patient care. Microsoft now plans clinical testing at Boston's Beth Israel Deaconess to see whether its AI can match these results in real-world settings.

Uber is rolling out ChatGPT Work to its roughly 30,000 employees worldwide, expanding its partnership with OpenAI after testing the enterprise AI tool with 1,000 staff earlier this year.

During the pilot, 80% of early users worked outside engineering. Teams used it to analyze operations data, localize training materials, and build personalized dashboards for daily work.

Uber hasn't settled on specific use cases or disclosed the rollout's price. Instead, it'll watch which workflows stick, making 30,000 employees part of a company-wide AI experiment.

Chinese humanoid robot makers, including UBTech, Unitree, and AgiBot, are expanding into South Korea, pursuing factory trials and commercial deployments as the country ramps up investment in physical AI.

Seoul has proposed $2.3 billion in physical AI funding for 2027, more than triple this year's level, aiming to modernize manufacturing as its working-age population shrinks and economic growth slows.

South Korea wants to lead physical AI, yet Chinese suppliers are already securing local trials. For now, the robotics race involves as much cooperation as competition on Korean factory floors.

A) Yes - 72% 🏆

B) No - 28%