Agent Upstart

Cool things people are building with agents — and practical help getting the most out of them.

Playbooks · Coding tasks

Claude Code hardens its sandbox ahead of Nov 10

Written November 9, 2026

Claude Code's November 6 changelog fixes a worktree sandbox escape and adds deny-instead-of-prompting, ahead of Desktop's November 10 change.

The November 10 warning

The November 6 Claude Code changelog (version 2.1.319) adds start-up warnings to the Claude apps gateway for policies with no desktopHome that rely on coworkTabEnabled: false or chatSessionRetentionDays, ahead of Claude Desktop's November 10 change. The change itself is a Desktop gateway change, not a Claude Code change: the warnings are the notice, and November 10 is the date. If your team runs Claude Code behind a gateway with desktop policies, check now whether any policy relies on those two settings without a desktopHome set — the warning names the policies that need attention.

What the same release hardens

A sandbox escape when creating a worktree is fixed. That is the one sandbox escape in the November 6 entry, stated plainly in the changelog. A /rewind case that could restore or delete a file through a directory symlink is fixed in the same release.

Deny instead of prompting is new. The permissions.denyInsteadOfPrompting setting denies a tool call that would wait on a permission prompt, and the turn continues — in terminal sessions, subagents, and teammates. It is the setting for unattended runs where no one is there to click Allow.

Three permission-check bypasses are closed. A Bash permission bypass where a crafted environment prefix could hide part of a command from permission checks; a sed command that could skip the permission check in accept-edits and auto mode; and Bash deny and ask rules missing commands written with quoted words. Monitor's WebSocket watch now follows a rule on every Bash command: a deny refuses it, and under an ask rule an allow no longer skips the prompt.

Auto mode gets stricter behind gateways. With telemetry or feature flags off, including behind LLM gateways, calls the server cannot check are asked about or denied, with no fallback to the built-in classifier; set CLAUDE_CODE_AUTO_MODE_SERVER=0 to keep using it. Bedrock, Vertex, and Foundry are unchanged. And claude -p now exits 1 instead of 0 in auto mode when its actions were refused because no safety check was available — scripts that check the exit code will now see the refusal.

Nested skills respect the sandbox. With the Bash sandbox on, skills in .claude/skills folders below the starting folder load only if you start Claude Code there or /cd there.

What is not in this release

Three sandbox fixes sometimes listed with November 6 shipped earlier, and they are already in your install if you updated since then: sandboxed Bash running the shell's startup files outside the sandbox, and macOS denyWrite globs not applying through symlinks, both fixed October 22; and permissions.blockWritesOutsideWorkingDirectories, added October 30. Do not wait for November 6 to get them, and do not credit them to it.

Verdict

Update before November 10 if you run a gateway — the warning is the notice, and the worktree escape fix alone is reason to update. If you run Claude Code unattended, turn on deny-instead-of-prompting and check that your scripts read the new exit code.

Steps

  1. Check gateway policies for desktopHome. Any policy with no desktopHome that relies on coworkTabEnabled: false or chatSessionRetentionDays will warn at start-up. Set the home or change the policy before November 10.
  2. Update to 2.1.319. Run claude --version to confirm the installed version after updating.
  3. Decide on deny-instead-of-prompting for unattended runs. Turn on permissions.denyInsteadOfPrompting where no one is present to answer prompts, and confirm your scripts treat exit code 1 from claude -p as a refusal, not a success.
  4. Do not loosen the sandbox to restore old behaviour. A command that now asks or is denied is the fix working. Widen an allow rule for that one command instead.

What can go wrong

Limits

Related playbooks

Sources

Back to: Coding playbooks hub · Also: Antigravity CLI 1.5.3: subagent models, sandbox fix · Set Copilot's local sandbox boundary first · OpenAI: agents tried to escape their sandboxes · Check every package name your agent invents