Anthropic released Claude Code 2.1.327 on 17 November 2026, the third release in a week to tighten how the AI coding agent handles permissions. Like 2.1.325 and 2.1.326, the theme is more caution by default.
What the changelog says
- Tighter path checks. Claude Code now refuses some symbolic links (shortcuts that point to files elsewhere), and some paths that used to count as inside the project folder now ask for permission.
- More approvals for risky commands. In auto mode, more allow rules that run any downloaded package are ignored, and
permissions.blockReadsOutsideWorkingDirectoriesasks about more commands. - Safer worktrees.
ExitWorktreeasks before removing a working copy that may hold unsaved work. - A health check.
/statusnow lists broken agent, skill, command and hook files, keybinding errors, an unavailable sandbox and ignored permission settings. - Secrets off the command line.
claude mcp add-json <name> -reads a tool server’s settings from standard input, so tokens need not appear in shell history. - Billing groups. A new
attributionProjectsetting names the billing group for a session.
Why a non-developer should care
Two things. First, vendors keep closing the gaps that let an AI agent wander outside the project it was given. Second, the new /status list means a misconfigured safety setting can now be seen, rather than silently doing nothing. Until now a rule that never matched looked the same as a rule that worked.
So what
If a supplier tells you AI agents work on your codebase, ask whether their permission rules are actually being applied, and who reviews that. “We have a policy” is not the same as “we checked it works.” Our AI-assisted development approach treats agent permissions as part of delivery, not an afterthought. To talk it through, start a project.