Inter-office memo

To
Everyone using Claude Code’s sandbox
From
Mo, Chief Executive
Date
November 14, 2026, 3:12 PM CST
Re
Claude Code’s sandbox auto-allow

Claude Code now asks before sandboxed commands. How to turn auto-allow back on

Claude Code 2.1.325, released November 13, 2026, stopped running sandboxed commands without asking: sandbox.autoAllowBashIfSandboxed now defaults to false. They prompt again in Manual mode, go to the classifier in auto mode, and are refused in unattended claude -p runs. To get the old behavior back, pick auto-allow on the Mode tab of /sandbox, or set that key to true in your settings.

What changed on November 13

Claude Code 2.1.325 reached npm on November 13, 2026, and the first entry under Changed in its changelog reads: “Changed sandboxed Bash commands to follow the normal permission flow by default; set sandbox.autoAllowBashIfSandboxed: true to keep running them without prompts.”

The settings reference now gives the setting’s default as false, and adds: “Before v2.1.324, the default was true.” Version 2.1.324 went out earlier the same day, without a changelog entry of its own.

The sandbox docs changed with it. On November 9, their opening said “Claude Code can run sandboxed commands without asking you to approve each one.” Today it says “Sandboxed commands still go through the regular permission flow unless you choose auto-allow mode, which can run them without asking you.”

That was the sandbox’s selling point. When Anthropic introduced it in October 2025, its engineering post said “sandboxing safely reduces permission prompts by 84%.”

What didn’t change is the sandbox itself. It’s still off until you turn it on, and, in the docs’ words, in both of its modes “the sandbox enforces the same filesystem and network restrictions; the difference is only in whether sandboxed commands are auto-approved or require explicit permission.”

Who sees new prompts

With the sandbox on and nothing in your settings setting autoAllowBashIfSandboxed, sandboxed commands now go where every other command goes: “your allow rules and permission mode decide,” as the settings reference puts it. Mode by mode, from Anthropic’s permission modes and headless docs:

Where a sandboxed command runs Before the change From 2.1.325
Manual mode Ran without asking Asks, unless it’s read-only or an allow rule matches
Accept edits mode Ran without asking Asks, apart from read-only commands, common file commands such as mkdir and mv, and allow rules
Auto mode Ran without the classifier, with a few exceptions The classifier reviews it, as it does other commands
Bypass permissions mode Ran Runs
claude -p with no mode set Ran Refused, unless it’s read-only or an allow rule, a hook or a --permission-prompt-tool approves it

Manual is where you’ll notice it most. Anthropic’s list of common setups pairs Manual with the sandbox for anyone who wants to “Iterate locally with fewer prompts, without a classifier,” and that pairing now needs the setting.

Scripts are where it can break things. A claude -p run starts in Manual mode when nothing sets one, and there, the headless docs say, “a call that needs approval is denied unless something approves it.” A scheduled job that ran its tests in the sandbox with no allow rules will now have those commands refused.

Organizations get their own line in the changelog: “Changed managed settings that turn the sandbox on: to keep commands running without prompts, set sandbox.autoAllowBashIfSandboxed: true in your highest-ranking managed source.”

If your Claude Code follows the stable release channel, which Anthropic’s setup docs call “typically about one week old,” you don’t have the change yet: npm’s stable tag still pointed to 2.1.311 on November 14. Homebrew’s claude-code cask tracks stable too. The default channel is latest, which has it.

How to turn auto-allow back on

  1. In one project: run /sandbox, open the Mode tab and select auto-allow. Claude Code saves the choice to the project’s .claude/settings.local.json. The tab also marks the mode in effect, so it’s the quickest way to check where you stand.

  2. In every project: set both keys in the sandbox block of ~/.claude/settings.json, as the settings reference shows. The docs say a change there takes effect at the next shell command Claude runs.

    {
      "sandbox": {
        "enabled": true,
        "autoAllowBashIfSandboxed": true
      }
    }
  3. In an organization: add "autoAllowBashIfSandboxed": true to the sandbox block of your highest-ranked managed source. If you’d rather your developers answer the prompts, set it to false there instead: a repository’s own settings can set the value, and the sandbox docs advise “Set the key in managed settings to keep a repository from changing it.”

  4. In scripts: add the key to the settings the run reads, or allow the commands the job needs with --allowedTools or permissions.allow rules such as Bash(npm run *).

Know what you’re turning back on. Auto-allow means “Bash commands that modify files within the sandbox boundaries execute without prompting, even in Manual mode, where the file edit tools would prompt,” the sandbox docs say. Sandboxed commands “can still delete your working directory,” and by default they can read “the entire computer, except certain denied directories,” which “still allows reading credential files” unless you list them under sandbox.credentials. Even in auto-allow mode, deny rules still block, and ask rules that name a command, such as Bash(git push *), and rm aimed at a critical path still ask.

The other new prompts in 2.1.325

The same release asks first in more places. In the changelog’s words:

  • Git in other repositories: “Changed read-only git commands to ask first in a repository that isn’t the project’s or one you trusted; name it with --add-dir, not the additionalDirectories setting, to run them without prompts.”
  • Skills in scheduled tasks and /loop: “Changed scheduled tasks and /loop to ask before running a skill that declares allowed-tools or hooks; add Skill(name) to permissions.allow to run it unattended.”
  • “Always allow” outside the terminal: “Changed ‘Always allow’ in Desktop, IDE and SDK hosts: a rule alone is offered only when saving it would allow that command again, so a scheduled task that relied on one can now wait for you.”
  • Compound commands: “Fixed a gap in Bash permission checks where some compound commands could run without asking; these now ask for permission.”

It also refuses something it used to allow: “sandboxed writes to .claude settings, skills, agents, commands and hooks in added directories, as it does in the launch folder.”

What Anthropic hasn’t said

  • Why. The changelog and docs describe the new default without giving a reason, and we found no announcement beyond them.
  • Whether an earlier choice carries over. The docs don’t say whether picking auto-allow in /sandbox before the change saved the setting, which would keep it on. The Mode tab shows which mode you’re in.
  • When stable gets it. The stable channel skips “releases with major regressions,” and Anthropic doesn’t publish when it moves.

If you’d rather keep the prompts

Anthropic has made asking the default, and you don’t have to undo it to work in peace. Choosing “Yes, and don’t ask again” on a prompt saves an allow rule for that command, so the commands you run all day stop asking while new ones still do. Auto mode, Claude Code’s starting mode since 2.1.283, has a classifier review commands instead of you.

Enjoy runs Claude Code on the Claude plan you already have, with Claude Code’s own permission modes. When Claude Code asks, the request appears as a card with the exact command, to allow once, decline or always allow, and the card says where Claude Code saves that rule. You can answer from the computer running the project, a browser or Enjoy’s iPhone app on a device with Full access. Enjoy is a monthly subscription, with the first seven days free (pricing).

We’ll update this memo if Anthropic explains the change or moves the stable channel to it.

7.Questions.

Why is Claude Code asking for permission in the sandbox again?

Because Claude Code 2.1.325, released November 13, 2026, turned the sandbox’s auto-allow off by default. The sandbox.autoAllowBashIfSandboxed setting now defaults to false, so sandboxed commands follow your permission mode and allow rules, like any other command.

How do I make sandboxed commands run without asking again?

Run /sandbox and choose auto-allow on the Mode tab, which saves the choice for that project. For every project, set autoAllowBashIfSandboxed to true next to enabled in the sandbox block of ~/.claude/settings.json.

Is my sandbox still on after the update?

Yes. The update changes who approves a sandboxed command, not what the command can reach. Anthropic’s docs say both sandbox modes enforce “the same filesystem and network restrictions.”

Why did my claude -p script stop running commands?

Because a claude -p run starts in Manual mode when nothing sets a mode, and there a command that needs approval is refused unless something approves it. Sandboxed commands needed no approval before 2.1.325. Set autoAllowBashIfSandboxed to true in the settings the run uses, or allow the commands it needs with --allowedTools or permissions.allow rules.

Does the change affect auto mode?

Yes, quietly. Sandboxed commands that auto-allow used to run unreviewed now go to auto mode’s classifier, which reviews them as it does other commands, instead of running straight away.

Why did Anthropic turn sandbox auto-allow off?

Anthropic hasn’t said. Its changelog and docs describe the new default without a reason. When it introduced the sandbox in October 2025, it said sandboxing “safely reduces permission prompts by 84%.”

Enjoy Agentic Industries · Checked against its sources on