ClaudeClaude ProThe paid subscription for the Claude AI assistant from Anthropic. It unlocks connectors to outside tools.Open the glossary Code 2.1.325 shipped on npm on 13 November 2026 (UTC) with 139 changelog entries. Upgrade if you rely on Bash permission rules or the sandbox, but first check 3 default changes that can make an unattended task stop and wait for approval.

Claude Code is the Anthropic coding agentAI agentAn AI program that performs work steps by itself, for example reading a message, drafting a reply, and recording the result.Open the glossary. It runs in the terminal, VS Code, the desktop app, and the cloud. The agent reads files, runs Bash commands, and calls MCPMCPA standard that lets an AI program use your tools and data under clear permissions.Open the glossary tools for you. Permission rules and the sandbox decide which commands run without a question.

What changed in Claude Code 2.1.325?

Release 2.1.325 closes several permission-rule gaps and changes how the sandbox asks for approval. In the official Claude Code changelog, 79 entries start with "Fixed", 28 start with "Changed", and 11 start with "Improved".

Version 2.1.324 shipped 85 minutes earlier on npm and has no separate notes. This release came 1 day after Claude Code 2.1.323, which closed 20 permission and sandbox gaps.

npm channelVersion on 16 November 2026What it means for you
latest2.1.325New installs and automatic updates get this version.
next2.1.325Same as latest.
stable2.1.311Does not include the 2.1.323 and 2.1.325 fixes.

Which security gaps does it close?

This release closes 1 zstd decoding security issue and several gaps that let a request get past Bash permission rules.

  • A security issue when Claude Code decodes zstd-compressed responses.
  • A Bash allow rule could run a different program than the one it names.
  • Some compound commands could run without a question.
  • Relative deny rules such as Read(./.env) stopped covering the project after a cd.
  • Deny rules did not apply to an approval that arrived after a restart in self-hosted runner and claude remote-control sessions.
  • The Bash sandbox did not fully protect the shared git folder of a linked worktree.
  • Background sessions on a claude.ai login used, and billed, an APIAPIThe official door 2 systems use to exchange data, without anybody copying it by hand.Open the glossary key from the shell that started the background service.
  • Token counting on Amazon Bedrock sent the conversation to Bedrock when the allowedProviders setting did not allow it.

The changelog gives no CVE number or score for these entries. Treat the permission-rule entries as security fixes, because deny rules are the main fence around an agent on your machine.

Which 3 default changes can stop an unattended task?

Three default changes make Claude Code ask for approval more often, also in tasks that run with nobody at the screen.

ChangeEffectHow to keep the old behavior
Sandboxed Bash commands follow the normal permission flowCommands that ran without a prompt now ask first.Set sandbox.autoAllowBashIfSandboxed: true. For managed settings, set it in the highest-ranking managed source.
Scheduled tasks and /loop ask before they run a skill that declares allowed-tools or hooksA daily routine can stop and wait for an answer.Add Skill(name) to permissions.allow.
Read-only git commands ask in a repository outside the projectAn agent that reads another repository now asks.Name that repository with --add-dir, not the additionalDirectories setting.
Amber toll gate barrier lowered in front of a car, a symbol of sandboxed Bash commands that now wait for approval
Since 2.1.325, a sandboxed Bash command stops at the approval barrier, like a car at a toll gate.

"Always allow" in Desktop, IDE, and SDK hosts also changed. Claude Code offers a rule alone only when that rule allows the same command again. A scheduled task that relied on such a rule can now wait for you.

How does it feel in daily use?

In interactive use, the most visible change is more permission prompts for sandboxed Bash commands.

Some parts feel better. When an organization policy blocks a request, the error now gives the next step: run /rewind and change the message, or run /clear. A subagent that hits the output token limit now gets a request to give the full answer again, more briefly. MCP tasks also tell Claude when a task waits for your answer.

Some parts add friction. Developers who let the sandbox run Bash without prompts must set it again. One macOS user reported .cc-writes and .claude folders in many projects (issue #110209, marked as a duplicate).

UserWhat to check
Business owner with daily routinesCheck that automatic reports still go out after the upgrade.
Marketer who uses skillsA skill with allowed-tools now needs an explicit allow rule in a scheduled task.
DeveloperReview Bash allow and deny rules, then choose the sandbox setting on purpose.

Example scenario (an illustration based on the changelog): a marketing team runs an ad report skill every morning as a scheduled task. The skill declares allowed-tools. After the upgrade, the task stops and waits for approval, so the report does not go out until a person opens the session. The fix: add Skill(ad-report) to permissions.allow, then run the task once by hand.

Conveyor belt with 1 amber box held at a closed sliding door, a symbol of a scheduled task that waits for skill approval
1 skill without an allow rule can hold a whole scheduled task until a person approves it.

Our check on 16 November 2026: our Linux workstation returned 2.1.325 (Claude Code) for claude --version, and this routine session ran on that version with no crash. This is 1 session, not a stability test.

Which early bugs have users reported?

In the first 2 days, users reported crashes on Linux and Windows that did not occur on 2.1.323.

IssueDateSymptom
#10989514 November 2026Linux x64: segfault in the JITWorker thread, 2 times in about 90 seconds. The reporter uses an AMD Zen 1 CPU. Labeled regression.
#11025915 November 2026Windows: crash at startup with exit code 0xC0000409. The reporter traced the trigger to 1 cached feature flag.
#11020915 November 2026macOS: .cc-writes and .claude folders appear in many projects.

All 3 issues were open on 16 November 2026. Each report comes from 1 machine, so it may not apply to yours.

Should you upgrade now?

Upgrade now if your security depends on Bash deny rules or the sandbox, because 2.1.325 closes gaps that let a request get past those rules.

Hold the upgrade on machines that run important routines until you set skill and sandbox permissions. Go back to 2.1.323 if your machine has a crash like issue #109895 or #110259. The stable channel is still 2.1.311, so teams on stable do not have these fixes yet.

What are the 6 steps for a safe upgrade?

Test 1 machine first, then check an interactive session and a scheduled task before you upgrade other machines.

  1. Run claude --version to check the version.
  2. Open settings.json and write down your Bash allow and deny rules.
  3. Upgrade 1 test machine, then run 1 interactive session and 1 scheduled task.
  4. If a task waits for approval, add Skill(name) to permissions.allow, or set sandbox.autoAllowBashIfSandboxed on purpose.
  5. If Claude Code crashes at startup, install 2.1.323 again. For an npm install: npm install -g @anthropic-ai/claude-code@2.1.323.
  6. Upgrade the other machines after the test machine runs 1 day with no problem.

If your agent uses MCP servers, also read the Model Context Protocol guide. This release fixes 6 MCP task and connection problems.

Conclusion

Claude Code 2.1.325 is worth installing because it closes Bash permission-rule gaps. Before you upgrade production machines, test 1 scheduled task and set skill permissions, so daily routines do not stop and wait for approval.

Do you want Claude Code routines to run safely with nobody watching? See our AI business process automation service or join the Claude class for marketing managers.

FAQ

Does Claude Code 2.1.325 fix security issues?

Yes. The changelog lists 1 zstd decoding security issue and several gaps that let a request get past Bash permission rules. It gives no CVE number.

Why did my scheduled task stop after the upgrade?

Since 2.1.325, scheduled tasks and /loop ask before they run a skill that declares allowed-tools or hooks. Add Skill(name) to permissions.allow so that skill runs unattended.

How do I get sandboxed Bash without prompts again?

Set sandbox.autoAllowBashIfSandboxed to true. For managed settings, set that value in the managed source with the highest rank.

Must I go back to an older version if Claude Code crashes?

Go back to 2.1.323 only if your machine crashes at startup, like report #109895 on Linux and #110259 on Windows. Watch both issues before you upgrade again.