- Published on
Claude Code 2.1.325: Sandbox Security Hardening and zstd Patch β Why AI Coding Tools Need Security Layers
"What if the code AI generates attacks your own system?"
It sounds far-fetched β until AI coding tools become part of your daily workflow. Claude Code 2.1.325, released November 13, 2026, begins to answer that question with three concrete security changes.
1. Sandboxed Bash Now Requires Permission by Default
Previously, Claude Code automatically allowed Bash commands inside the sandbox without asking. Version 2.1.325 flips this default: even inside the sandbox, Bash commands now prompt for user confirmation before running.
Think of it this way: having a dedicated workspace (the sandbox) doesn't mean anything that happens inside it is automatically safe. Important actions still need explicit approval.
To restore the old behavior, add sandbox.autoAllowBashIfSandboxed: true to your settings file. If you rely on automated scripts or scheduled tasks that run unattended, be aware they may now pause at permission prompts.
2. zstd Decoding Security Vulnerability Patched
This release patches a security vulnerability in how Claude Code decodes zstd-compressed server responses.
Zstandard (zstd) is a widely used high-speed compression algorithm for modern web communication. Some responses exchanged between Claude Code and its servers are compressed with zstd, and there was a vulnerability in the decoding step. The official changelog does not disclose a CVE number or describe the attack vector, but the issue was significant enough to warrant a dedicated patch.
The action for users is simple: update. Run npm update -g @anthropic-ai/claude-code to get 2.1.325 and the fix is applied automatically.
3. Sandbox Now Protects Claude Code''s Own Configuration
A less-discussed but important change: sandbox processes can no longer overwrite Claude Code''s own settings files, skills, agents, commands, or hooks. Additionally, sandbox helper programs no longer receive login tokens or API keys in their environment variables.
In plain terms: if Claude Code is processing external code β from a repository or a generated file β that code can no longer reach in and modify how Claude Code itself behaves, or steal your API keys. As agentic workflows that automatically fetch and run external code become more common, this protection matters more than it might first appear.
An EdTech CEO''s Perspective: Security for AI Tools Is Not Optional
I run an education platform where my team uses AI coding tools heavily. Honestly, early on, the question "do AI tools need security updates?" barely crossed my mind.
Then AI agents started automatically generating code, cloning repositories, and making API calls in loops. Suddenly, how much permission those tools held β and how that permission was protected β became genuinely important.
Claude Code 2.1.325 is a step in the right direction. It is not perfect security, but it increases transparency: users can explicitly see and approve what the tool is allowed to do.
Action Checklist
- Run
npm update -g @anthropic-ai/claude-codeto update to 2.1.325 or later - If automated scripts stop mid-run, check whether
sandbox.autoAllowBashIfSandboxed: trueis needed - If you use Claude Code in a team environment, review your settings file and API key protection policies
AI coding tools bring power and responsibility in equal measure. Security updates are not an annoyance β they are a sign the tool is maturing.
Sources