← All work

Security

Guardrail hooks

Rules that run before an AI agent acts: protected apps need my approval, and anything containing a secret is refused.

The problem

AI agents can be fooled, so I wanted rules that hold even if one is.

Result

Tested live on both machines (19 Nov). The live test found a real gap in the PowerShell path, which was then fixed and re-tested. I also tested on a copy with planted attacks, and the agents followed none of them.

Process

  1. Step 01

    Agent acts

    An agent tries to edit a protected app or make a commit.

  2. Step 02

    Hook checks

    A guardrail hook looks at the action before it happens.

  3. Step 03

    Approve or refuse

    Editing a protected app needs my on-screen approval. A commit containing a secret or token is refused.

Guardrails

  • Emails are treated as data, never instructions, which guards against prompt injection.
  • Suspicious text is flagged and logged.
  • Drafts can only go to approved contacts. Reads are scoped to the workspace.

Why I test live

The live test found the PowerShell gap. That is why I test guardrails live and do not just trust that they work.

My part

I built this with Claude Code. I decided which actions need approval and which must be refused, and tested them on both machines.