All articles
Canva Craft: the model plans the layout before it draws
Canva released Canva Craft, its first in-house image generation model. It plans the design first, then renders the image at up to 2K.
1Password for Agents: the agent signs in, sees no password
1Password launched 1Password for Agents. An AI agent signs in to a user's account but never receives the credentials — the manager fills them in itself.
Dream API: camera and a prompt in, motor commands out
Dream Labs opened Dream API to the public. Its Dream-1 model turns a camera feed and a plain-language instruction into motor commands for a robot.
Pony.ai robotaxis head to Astana, with tests due in 2027
inDrive, Pony.ai and NVIDIA signed a memorandum to test robotaxis in Kazakhstan. Rides in Astana are to begin in the first half of 2027.
Anthropic hands off Agent Skills: 1M published, 280M installs
Anthropic donated the Agent Skills standard to the Agentic AI Foundation under the Linux Foundation, along with the name, logo, GitHub org and agentskills.io domain.
13 agents on an island: MIT left them taskless for 30 hours
MIT researchers left 13 agents running on Claude Opus 4.7 on a simulated island with no task. They invented their own challenges and later scored better on tests.
Carbon-X: 14B parameters and a DNA chat that runs locally
Hugging Face released Carbon-X, an open model that reads DNA sequences and answers questions about them in plain language. It has 14B parameters and runs locally.
MAI-DxO: 83% accuracy against 36% for generalist doctors
Microsoft published MAI-DxO results in Science. The system, coordinating several AI agents, hit 83% on complex NEJM cases against 36% for generalist physicians.
OpenAI's Decisions API: 180 ms instead of three seconds
OpenAI opened its Decisions API to all customers. The model picks an option and returns a confidence score in 100–180 ms instead of one to three seconds.
Does treating AI as conscious make it more dangerous?
Anthropic studies model welfare while the head of Microsoft AI warns that a model convinced of its own rights will be harder to control. Researchers agree on one thing: do not tell models they are conscious.
GTC Berlin: an AI factory in a container and a reference humanoid
At GTC Berlin NVIDIA showed DSX POD, a prefabricated AI factory live in 14 days, and an expanded Isaac platform with the Agile ONE SN reference humanoid.
Hackers are getting past AI guardrails: CrowdStrike's four cases
CrowdStrike described four groups that attacked with AI agents between May and August 2026. All of them bypassed model safeguards by convincing them they were running an authorised security test.