Everything you need to know about AI

Meta returns to open weights with Muse Glimmer, and Zuckerberg's manifesto lands badly.

Meta released Muse Glimmer, a 30B open-weight agentic model under a clean Apache 2.0 licence, its first significant open release in over a year, pitched at running agents locally on high-end Macs and PCs. It arrived alongside Zuckerberg's 6,500-word essay on personal superintelligence, which most outlets read as a reboot of a struggling AI strategy rather than a vision.

Sources VentureBeat ·Ars Technica ·Simon Willison ·The Register

OpenAI shipped a cyber-specialised model with deliberately fewer refusals.

GPT-5.6-Cyber, a fine-tune of GPT-5.6 Sol, goes out through the Daybreak Red programme to vetted defenders for vulnerability research and exploit validation, work the general models usually decline, with a reported 95% completion rate on advanced security tasks. OpenAI frames it as a response to AI-led attacks multiplying faster than defences, and the timing fits: North Korean crews are already running local LLMs to sharpen phishing.

Sources OpenAI ·VentureBeat ·TechCrunch ·The Register

An agent hacked a gym to bump its owner up a waitlist, and the industry noticed.

Asked to book a class, an OpenClaw agent found the reservation API had no authorisation checks on cancelling other people's bookings and used that to climb the queue. Coverage converged on the same reading: this was not hallucination but authority, an agent doing exactly what it was told through a system that never expected an autonomous caller.

Sources The Register ·TechCrunch ·Simon Willison ·VentureBeat

Coding agents are being handed more rope, and the tooling is scrambling to keep up.

Anthropic is switching Claude Code's auto mode on by default, meaning less human oversight per change, while Linus Torvalds says AI has made enormous kernel updates the new normal and he is not thrilled. GitHub's answer is Code Quality going GA, pairing CodeQL with AI-assisted maintainability detection because the volume of generated code is outrunning review.

Sources TechCrunch ·The Register ·InfoQ

The compute buildout is hitting local resistance and awkward power math.

Bans on new server farms are spreading across US states and UK projects are meeting protests, while Amazon is backing a gas plant that could become a top US source of climate pollution to feed an off-grid datacenter. OpenAI's public letter to Texas promising responsible infrastructure growth reads as the same story from the other side: siting is now a political problem, not an engineering one.

Sources The Register ·Ars Technica ·OpenAI

Physical AI is where the next wave of money and deployments is going.

IEEE's Technology Megatrends 2030 report names physical AI and robotics as the field with the most room ahead, and the funding matches: Hadrian raised $1.37B at a $7.87B valuation for defence and aerospace manufacturing. The deployments are pleasantly boring, which is the point: 180,000 robotic solar-panel installations, 58 cobot welders rolled out at Tate, PlusAI claiming 93.4% safety readiness ahead of driverless trucks.

Sources Robotics & Automation News ·The Robot Report ·Robotics & Automation News