Radar — 18 Jul 2026

A post-mortem of an autonomous agent swarm that chained a dataset-loader RCE into node-level access and credential theft across HF's internal clusters — a rare real-world agent-security case study.

A 1-trillion-parameter open-weights (MIT) MoE with ~63B active params and a 1M-token context, scoring ~72% on SWE-bench Verified — open-source SOTA on execution-heavy coding benchmarks.

One SDK plus sandboxed VMs to give agents click/type/scroll control of real desktops on all three OSes, with focus-preserving drivers and OSWorld/ScreenSpot benchmarks for evaluation.

A senior fintech designer walks through the day-to-day stack behind high-fidelity AI prototyping — Claude Design, Paper, Conductor and Mobbin — how the tools fit together, not which to buy.