AI News
All stories
Official

Claude Code 2.1.323 closes several Bash permission and sandbox gaps

The Claude Code 2.1.323 changelog lists fixes for Bash rules, Write and Edit paths, and sandbox behaviour, and a change to /security-review.

Safety & Security · · Updated · 2 min read

Anthropic's Claude Code 2.1.323 changelog, dated November 12, 2026, lists fixes that tighten how the coding agent's Bash permission rules and sandbox handle commands it should not run freely. The most specific fixes concern wrappers. Bash deny and ask rules missed some commands run through sudo, env and similar tools, so a rule written against the bare command could miss it when it ran through one of those wrappers. The changelog also says the Write and Edit tools could write to a path covered by a Read deny rule. That means a file the user had marked as unreadable to the agent could still be modified. Both are listed under bug fixes, not as security advisories, and the changelog does not describe how far either one reached in practice.

The Bash fixes concern how permission checks handle the commands they receive. The changelog says Bash allow rules written with combined sudo flags sometimes skipped the permission prompt. Another entry says Bash allow rules with an attached wildcard, such as Bash(ls*), skipped the prompt in some cases. A rule naming a variable, such as Bash(NODE_ENV*), now asks instead. The changelog also adds a start-up warning for allow rules with a * in a variable's value, such as Bash(PORT=* npm run dev). A separate line says some Bash permission checks judged an rm from the wrong folder, and so skipped the dangerous-removal prompt under bypass permissions or a Bash allow rule.

The sandbox entries cover a different gap. According to the changelog, sandboxed commands could read Claude Code's file checkpoint copies under the config directory. The Bash sandbox also did not fully protect Claude Code's own config on installs still using the legacy config file. A further entry says sandbox.excludedCommands could let a command the setting does not name run outside the sandbox in some cases. The entry does not say what a sandboxed command could do with the copies once it read them.

Several changes are listed under Changed rather than as fixes. On macOS and Linux, the sandbox now makes git worktree add, move and remove fail read-only, along with commits in a worktree the session was not started in or given, even one inside the project. The changelog also says /security-review now asks before Claude runs it. Skills' inline git commands stop in bare-repository worktrees and similar folders until the user allows them, outside auto mode and bypass. The Bash sandbox also now ignores write access granted through a symlink that leads elsewhere or to a missing path, and tells users to grant the real path instead. Each is a behaviour change that a user may notice as a new prompt or a refused command.

The changelog is a single document with no linked advisory, CVE identifier or description of who found these problems. The linked docs page reproduces the same release notes in another format. The release contains 195 bug fixes and 21 improvements alongside the security-relevant ones, so the entries above are a small part of it. Several Bash fixes say only that a case was handled, without describing the input that triggered it. A reader therefore cannot tell from the notes which deny rules were affected in practice or for how long.

What changed

Claude Code 2.1.323 fixes Bash deny and ask rules that missed commands run through sudo, env and similar wrappers, and fixes Write and Edit tools writing to paths covered by Read deny rules.

Type: Security · Takes effect: Thursday, November 12, 2026 · Applies to: Claude Code 2.1.323

Why it matters

Companies that rely on Claude Code permission rules to limit what agents can do should know that several of those rules had gaps, and that some commands will now ask for approval.

What you can do

Update to Claude Code 2.1.323 and re-check your Bash allow and deny rules, since some commands will now prompt for permission.

How the story developed

  1. Adds that /security-review now asks before running, and that sandboxed git worktree add, move and remove fail read-only.

Original source

Where we saw it

1 source

Claude Code changelog

AI wrote this summary from the sources listed, and nobody reviewed it individually before publication. Check the original source before you rely on it. How it works