Malicious package protection for AI agents
Known malware is blocked before your agent installs it. Brand-new packages, brand-new versions and names the registry has never heard of are flagged.
● Bash(npm install @acme/pay-utils)
⎿Muzzle blocked this: @acme/pay-utils (npm) is known malware (MAL-2026-0412). To allow it, click Allow on this entry in Muzzle's Activity.
New package npm install react-form-kitt · was first published 3 days ago, and runs an install script
Unknown package pip install fastapi-auth-helpers · isn't on PyPI
What gets blocked or flagged
Known malware
BlockA brand-new package
WarnA brand-new version
WarnA name the registry doesn't have
WarnWarnings tell you if it runs an install script.
Works with
Packages named in the command:
- npm
- pnpm
- yarn
- bun
- npx / bunx / pnpm dlx
- pip
- python -m pip
- uv
- cargo add / install
- gem install
Installs from your own registry (--registry, --index-url) aren't looked up. RubyGems gets the malware check only.
Questions
Does it check my package.json?
No, only packages named in the install command, such as npm install zod or pip install requests.
What does the lookup send?
Only the package name, to OSV and the package's registry.
A package I trust is flagged.
Click Allow on its entry in Muzzle's Activity.
Related
Put Muzzle between your agents and the internet
Download for Mac
Free · Apple Silicon · macOS 12 or later