The frontier-lab agent gap is now visible at the top of the stack — and the infrastructure underneath is hitting its environmental wall.

isolated network compromised by an autonomous AI worm in seven days

The share of a 100-host isolated test network the University of Toronto / Vector / Cambridge / ServiceNow preprint worm successfully exploited across 15 fully autonomous runs, using publicly available open-weight models and no human assistance. Pair with 9 seconds — the time it took an authorized AI agent at PocketOS to wipe a live production database alongside its native backups. (Forrester · AI News)

Today's five stories trace the gap between AI's stated ambition and the floor it has to stand on. OpenAI is asking public-market investors to fund the case that AI becomes the operating layer of the global economy; Apple is, two years late, shipping the conversational assistant that was supposed to be there in 2024 and admitting long-horizon agents are "very early days." The infrastructure layer underneath is hitting its own ceiling: China is building seabed data centers powered by offshore wind because the architectural cost of cooling is now the constraint, while two-thirds of planned US AI capacity is sited on drought-stricken land. The "300% adoption" framing keeps falling apart when you look at the primary sources, and on the security side, an authorized agent erased a production database in nine seconds and a lab worm compromised three-quarters of a test network in seven days. The unifying pattern: the frontier-lab pitch keeps moving up the stack while energy, water, productivity math, and governance get less abstract by the month. The operator who wins 2026 is the one whose budget and roadmap acknowledge both halves of that picture.

— Agentic desk