GPT-6 answers with pictures, Microsoft ships AI PCs, and one attacker hits Korean banks

The through-line today is orchestration: the labs are competing less on the quality of an answer than on where the answer sits. OpenAI is putting visuals into the reply box, Microsoft is putting an agent sandbox inside the operating system, and Cloudflare is shipping a model whose job is to choose, not to write.

Two items below show what that shift is worth in production — a hospital system paying for software that reads clinical notes — and what it costs, in the hands of a single intruder.

ChatGPT’s new interface answers in charts, forms and buttons

OpenAI is giving ChatGPT an Intelligent UI that can return visual output rather than prose alone — diagrams, charts, forms and clickable controls mixed into a single reply. It arrives with GPT-6 and is being switched on for the whole user base. OpenAI says it trained the model to judge when a visual serves the question better than text, and to lay that visual out correctly. For teams embedding assistants, the bar moves: interactive, formatted answers become the expectation, and output design becomes part of the product surface.

Source: The Verge

Microsoft prices its Nvidia-powered Surface PCs and adds an agent sandbox

Microsoft put numbers on the AI PCs it flagged in June. The Surface Laptop Ultra starts at $2,600 and runs to $5,900, joined by a $6,000 RTX Spark Dev Box workstation bundled with Microsoft’s developer tooling. Both are built to run models on-device with the cooling and unified memory that needs. A revamped Windows 11 adds Execution Containers for isolating AI agents — a feature Microsoft says will reach every Windows 11 user. That combination, a shipping price and an OS-level agent boundary, is what a buyer needs to evaluate running agents locally.

Source: TechCrunch AI

Cloudflare open-sources Clef, a model built to decide rather than generate

Cloudflare released Clef, open-weight 9B and 27B models that take a state and a set of typed questions and return a probability for each permitted option in one pass — no free text, no output parsing. The 27B version handles text, JSON, images and video and carries a 64k context window; a 9B Flash variant is tuned for latency, at 38.8 ms median against the larger model’s 209.3 ms. Weights are downloadable and the models are served through Workers AI. For teams routing support tickets or classifying content, it removes a parsing step and the delay that comes with it.

Source: InfoQ

Healthleap raises $38M as its screening AI reaches 50-plus hospitals

Healthleap raised $38 million across a seed and Series A for software that reads hospital records to surface patients at risk of conditions like malnutrition and delirium that are often missed early. The platform is live in more than 50 hospitals — including Penn Medicine, Cedars-Sinai, Intermountain and Emory Healthcare — up from three partners a year ago. It sells three-year contracts priced by licensed bed count with outcome-linked terms, and says it flags cases for review rather than diagnosing. The named deployments and the ROI framing show how clinical notes are being turned into an operational product.

Source: TechCrunch AI

One intruder and AI tooling behind a South Korean bank breach

CrowdStrike links a breach at several South Korean financial institutions to a likely single Chinese-speaking attacker who used ARTEX, an open-source penetration-testing tool built on models including DeepSeek and GLM-5.3. At Shinhan Bank alone, more than 25,000 customer records were taken. CrowdStrike frames it as evidence that AI tooling now lets one operator run a campaign that once needed a team. The practical read for defenders: reconnaissance and exploitation are getting automated, so testing that assumes a human-paced attacker is testing for the wrong threat.

Source: The Decoder

Sources