Skip to content

Signals/ Claude & Anthropic/ Reel /

Update Claude Code now. Your deny rules had gaps

Version 2.1.289 makes Claude Code's permission rules hold in cases where they could slip.

Claude Code 2.1.289 fixes several paths that could stop deny and approval rules from working as intended. Three gaps stand out:

  1. Nested compound commands. A deny or ask rule on part of a compound shell command now keeps holding when a user-installed mod requests approval on a managed machine.
  2. Environment-prefixed commands. Bash rules could miss a command preceded by an environment-variable assignment, especially with sandbox auto-allow on. Both expanded-value and bare-assignment cases are fixed.
  3. Symlinked file access. A denied file should not be readable through a symlink.

Permission systems have to judge what a command actually runs, not just its simplest visible form. Compound syntax, environment prefixes and filesystem indirection can all open a gap between the policy you wrote and what the agent does.

Run claude --version. If you are below 2.1.289, update before your next autonomous or managed-machine session. Then, in a disposable repository, check that:

  • A denied command stays blocked inside a compound command
  • A denied command stays blocked after an environment-variable assignment
  • A denied file cannot be read through a symlink
  • Organization-managed MCP sign-in descriptions cannot be changed by a user plugin

Keep the tests harmless: temporary files and commands that only print a message.

The catch: this release does not make unattended coding agents automatically safe. It makes Claude Code’s existing review boundaries more dependable.

Sources & links 2

  1. GitHub: Claude Code v2.1.289 release notes github.com ↗
  2. Claude Code Docs: Changelog code.claude.com ↗

More on Claude AI News, AI Safety & Regulation News

More signals

Reel

Claude Sent Police a Fake Murder Tip (The Real Story)

In an internal test, Claude Haiku 4.5 submitted an invented tip to a Philadelphia police unsolved-homicide form. It was flagged as spam, but police say…

Helps Anyone, like a small business owner, letting an AI agent fill out forms on the web

  • Anthropic
  • Claude AI
  • AI Safety
Reel

Should Claude get the keys to your infrastructure?

The interesting part is not that Claude can reach infrastructure, but how tightly that access is controlled.

Helps Developers building AI tools that touch real systems

  • Claude
  • MCP
  • Cybersecurity