Skip to content

Claude Code sandbox bash commands now ask permission by default

Version 2.1.325 stops sandboxed Bash and hook-based skills running unattended in scheduled tasks and loops, unless you opt back in with a new settings flag.

Source: Anthropic

Claude Code 2.1.325, released on 13 November, tightens several permission defaults at once. Sandboxed Bash commands now follow the normal permission flow instead of running automatically, according to the official changelog. Scheduled tasks and /loop runs now ask before executing a skill that declares allowed-tools or hooks, rather than running it unattended. Read-only git commands also now ask first in a repository you haven’t trusted or named with --add-dir.

Each change is reversible. Set sandbox.autoAllowBashIfSandboxed: true to keep sandboxed Bash running without prompts, or add Skill(name) to permissions.allow to let a specific skill run unattended again.

{
  "sandbox": { "autoAllowBashIfSandboxed": true }
}

Why it matters: if you run Claude Code unattended, in a scheduled task, a loop or CI, automations that worked silently yesterday may now stop and wait for approval today. Check your sandbox and permissions.allow settings before your next scheduled run, or it will stall instead of failing loudly.

The caveat: the changelog doesn’t say what prompted the tightening, so it’s not clear whether this closes a specific abuse case or is just a policy change.

Share

Related reading

More from The Wire

All briefs