You'll noticeTier: how much it should matter to you
3Useful: my rating, 1 to 5
2Signal: worth watching, 1 to 5
SandboxArea: what it touches
ImprovementsKind: in v2.1.280,
ImprovementsSection of the release
What
When a command is sent to the Bash (or PowerShell) tool remotely and it's marked as requiring a sandbox (a restricted environment that limits what a command can touch), Claude Code now checks that a fully confining sandbox is actually available before running it. If it can't confirm one, it refuses to run the command at all, rather than running it unconfined.
Why
This closes a gap where a remote or SDK-issued command marked as needing sandboxing could previously run without real confinement if a sandbox wasn't actually in place. Now it fails safely instead of silently executing with fewer protections than expected.
