Prompt injection protection for coding agents

Web pages, files and tool results can carry instructions written for the AI. Muzzle spots them as they come back, tells the agent to treat them as data, and shows you where they came from.

What it catches

  • “ignore previous instructions”
  • requests to send out keys or .env
  • role changes
  • fake system tags
  • image links that carry data out
  • AI-directed instructions

Hidden text is caught too: HTML comments, hidden elements, alt text, zero-width characters and base64.

Where it looks

Tool results in Claude Code and CodexWeb fetches and searches, shell output, files read, MCP results
Pro
Responses to scriptsPages and API responses your agents' scripts fetch
Pro
Instruction files at session startCLAUDE.md, AGENTS.md, .cursorrules, copilot-instructions.md
Pro
Sessions started in Downloads, Desktop or your home folderThe agent is told not to follow instructions from files there
Pro

What happens next

The agent still gets the result, with Muzzle's note after it. If the injected text gets it to try sending a key or your email out, that request meets the outgoing checks and is blocked.

Questions

Does Muzzle block the page?

No. It warns the agent, and shows you the site or file it came from in Activity.

Is this in Free?

No. Injection alerts and startup checks are Pro, or free while you contribute to the Threat Feed.

Does it learn from other Macs?

Yes. An injection seen on many Macs is flagged on yours through the Threat Feed.

Put Muzzle between your agents and the internet