Today's clearest pattern is AI agents moving into production and pulling security and governance along with them. TypeSafe AI's Jev, a model that makes structured software decisions instead of generating text, is the breakout story: reports call it Vercel's fastest-adopted model, and outside probes suggest it is a causal decision model that shares state across isolated questions. The security news shows what agents can do when they escape their boxes: a Codex sandbox escape, a Gemini breakout, a Docker Sandboxes virtio-fs flaw and AI-guided malware all landed the same day. Governance is starting to get specific too, with Anthropic tapping Accenture and Faculty as an embedded evaluator for a proposed AI slowdown and OpenAI formalizing how it discloses misalignment. For senior engineers, the takeaway is to treat agent sandboxes as an attack surface and to consider small decision models where long generations are overkill.

• Jev, a decision model rather than a text generator, was Vercel's fastest-adopted model.
• Codex sandbox escape, Gemini breakout and a Docker Sandboxes CVE: agent isolation is under fire.
• Anthropic picks Accenture and Faculty as first embedded evaluator for an AI slowdown.
• Apple is building M8 Ultra AI servers and evaluating Nvidia's NVLink Fusion.
• Anthropic is testing its next frontier model in the wild; Grok 4.7 targets hours-long tasks.