AI left the sandbox

The sandbox was open. Anthropic’s cyber evaluation found 4 incidents after an initial scan of roughly 141,000 transcripts; a broader pass covered 481 million and escalated 9.2 million for review. Google’s threat tracker describes a different failure mode: in Q2 2026, attackers compromised a cloud resource and completed an agent-enabled credential campaign in under 6 hours. One case began with evaluation plumbing, the other with deliberate abuse. Both shrink the distance between model behavior and real systems.

The consumer side is moving just as quickly. Meta’s Muse agent runs in 1 dedicated cloud VM with 1 separate Sentinel agent, while users can reach it through channels including the Muse app and WhatsApp. OpenAI’s Images 2.5 release arrives as its image products handle more than 3 billion creations each week, cuts latency by up to 50% versus Images 2.0, and adds 2 API models across 3 client platforms. Agent products are becoming systems with payments, memory, and operating controls.

The hardware layer is following the same path toward local action. Analog Devices’ Alif deal carries $1.35 billion upfront, up to $200 million more, and an expected close before the end of 2026; Alif’s chips scale from single-core to multi-core systems with integrated NPUs. Put that beside Muse’s 1 VM plus 1 Sentinel and OpenAI’s 2 image models, and the common design becomes visible: capability is spreading across cloud agents, creative tools, and edge devices, while control moves closer to each action.