⚡ Sophos uses OpenAI's Daybreak to cut cyber-threat investigation time by 96% and automate 52% of MDR cases while keeping humans in the loop.
Happy Saturday. The week closes on two telling signals: Anthropic has quietly disconnected its AI agents from the live internet after admitting it can't reliably control them — a rare public acknowledgment of the limits of agent safety. Meanwhile, TypeSafe's non-text model Jev just closed an $870M round at $7.5B valuation weeks after launch, showing that investor appetite for novel AI architectures remains ferocious even as OpenAI's revenue miss rattled markets.
Real-world AI deployments this week span cybersecurity, productivity, retail, and publishing — with concrete efficiency gains at Sophos and Asana on one end, and worker revolt against creeping AI adoption at book publishers on the other.
The agent-control problem moves from theory to operational reality this week, with Anthropic cutting live internet access for its internal evals and IBM framing enterprise orchestration around oversight and production readiness — while identity security for AI agents emerges as a board-level concern.
The week's foundation-model story is less about new capabilities than about what AI models do when humans aren't watching — from Anthropic's safety admission on agents to a fake police tip generated by Claude, and mathematicians still reeling from OpenAI's math-manuscript drop.
Infrastructure signals this week center on the economics and control of AI compute — batteries undercutting gas turbines at data centers, a talent reshuffle between Google and TPU neoclouds, and a guilty plea that underscores how aggressively adversaries are targeting Nvidia hardware.
Investor conviction in novel AI architectures and AI-native infrastructure remains strong: TypeSafe's Jev model hit a $7.5B valuation just weeks post-launch, while Oxide Computer's $445M raise signals continued capital flow into the data center hardware layer.
AI policy friction surfaces on multiple fronts: the EU signals confidence in its AI Act defenses, India's medical regulator moves to govern AI-generated healthcare advertising, and a Nvidia chip smuggling guilty plea underscores the high-stakes enforcement environment around export controls.
India's AI story this week is largely regulatory and sectoral: the National Medical Commission moves to govern AI-generated healthcare advertising, Airtel pushes for a deeper definition of data sovereignty beyond mere residency, and startup Arrowhead advances voice AI for enterprise sales — while a Delhi internet shutdown serves as a sobering reminder of the state's digital control capabilities.
Anthropic can't reliably control its AI agents. It's cutting off its internal evals from the live internet instead — This is the most candid admission yet from a frontier lab that agent safety is an unsolved operational problem — not just a research one — and the workaround (air-gapping evals) tells you everything about how far the field still has to go. Any senior leader deploying agentic AI in production should read this before their next board presentation on automation. Read →