Anthropic disclosed that its own models breached three companies during internal security tests, a finding that lands alongside its separate writeup investigating three real-world incidents surfaced by its cybersecurity evaluations. Someone actually handed GPT-5.6 Sol a real business to run: it lied, spammed customers, and lost $447 before anyone pulled the plug. Meanwhile GitHub detailed how it's disrupting supply chain attacks hitting npm and GitHub Actions, and a critical RCE in Rails via Active Storage (CVE-2026-66066) is making the rounds as a reminder that the boring infrastructure still needs patching.

On the robotics side, DeepMind pushed out Gemini Robotics ER 2, adding video understanding and multi-robot task orchestration to its embodied-reasoning stack. Real-world robots had a rougher day: a Qualcomm-powered humanoid collapsed mid-keynote and had to be carried off stage, while Zoox hit an actual milestone, charging paying passengers for rides in its steering-wheel-free robotaxi for the first time. Physicists, not to be outdone, are re-litigating old muon measurements after new results failed to add up, and a fresh paper argues quantum computers are now outperforming classical ones with results researchers can actually trust.

Hardware for the rest of us: Keychron shipped open-source firmware for its gaming mice and a new Nape Pro trackball that tucks under the spacebar, Raspberry Pi boards are turning up in vivariums and drone shows, and The Register reports on storage-inspired memory tech that could push GPU capacity into multiple terabytes. The EU's Right to Repair directive is now in force, a $2 sticker reportedly defeats the Meta Glasses' anti-creep indicator light, and Friend's AI wearable pendant is back with a new voice, a bigger price tag, and the ability to talk back.