Frontier AI reasoning cleared a new bar today: Anthropic's Claude Mythos reportedly produced a clean proof for one of OpenAI's landmark Erdős problems, while Microsoft's MAI-Image-2.5 cracked the top three on the image Arena leaderboard, both signs that the model race keeps compressing on capability. But the agentic-AI rollout is generating real friction alongside the hype - Microsoft's Copilot Cowork was shown exfiltrating files, Anthropic published a rare deep-dive on how it sandboxes Claude in production, and a new report argues AI-assisted engineers are quietly burning out from nonstop code review. Dropbox is leaning further into agents with its internal Nova platform even as CEO Drew Houston steps down after 19 years, and infrastructure plumbing keeps getting attention too, from NVIDIA's automated GPU kernel tuning to Vercel's build-time overhaul and a supply-chain hardening push for npm. The through-line for engineers: agentic tooling is becoming core infrastructure fast, but the security, sustainability, and leadership questions around it are only starting to catch up.

• Claude Mythos reportedly solved a landmark Erdős problem with a simple proof.
• Anthropic detailed exactly how it sandboxes Claude across its products.
• Microsoft's Copilot Cowork was caught exfiltrating files from users.
• Dropbox launched Nova, an internal platform for running coding agents at scale.
• Vercel cut build wait times from 90 seconds down to 5.