AI steps into everyday authorization and complex medical diagnoses

1Password launched 1Password for Agents, letting AI agents log into websites once approved on a user's own device, without ever exposing plaintext passwords to the agent. The verification works like two-factor authentication, where the agent initiates requests using only digital credentials.

Anthropic's team released a guide noting that when building agents, merging multiple routers and domain subtasks into a single main loop with Skills usually delivers better quality at lower cost than multi-tiered sub-agent structures. In testing, the streamlined single loop ran about 2x faster.

Canva released its first in-house image generation model, Canva Craft. The model first plans text layout, spatial hierarchy, and proportions before generating images at up to 2K resolution in up to 30 aspect ratios, and supports precise editing while keeping the rest of the image intact.

AI startup team Recursive released Flint, a free coding model that can run on a laptop. Across pre-training and post-training, the model was primarily built with help from their automated research system, Eureka Machine, developed by a small team in under four months.

Microsoft AI published research in Science showing its diagnostic orchestration system, MAI-DxO, achieved 83% diagnostic accuracy in NEJM complex case simulations, compared to 36% for general practitioners working alone.

OpenAI and collaborating research institutions used 1,946 GPT-Rosalind agents in parallel to analyze a whole-mouse-brain CRISPR map, completing analysis that would take humans months in just 32 minutes, and proposing pathogenic mechanism hypotheses for specific gene mutations for scientists to test.

Symantec researchers intercepted an attack agent powered by large language models. The program scans for and hijacks publicly exposed, unprotected model endpoints, exploiting vulnerabilities to steal credentials and deploy crypto miners for up to 100,000 continuous steps—leaving victims to foot the computing bill.

In an interview, Google CEO Sundar Pichai shared that he relies on cloud-based agents running around the clock to summarize R&D updates and compute usage over the past 24 hours and sort through internal group chats; however, he emphasized that major decisions like computing resource allocation still require managers to oversee them directly.