ZT4AI Self-Audit

Audit your agent's skills, workspace, and configuration against Zero Trust for AI principles.

Background

AI agents process instructions and data as indistinguishable tokens in a context window. This means:

• Skill files loaded into context can inject behavioral instructions
• Workspace files (SOUL.md, AGENTS.md) are both operating instructions AND attack surface
• External inputs (web content, emails, ClawHub skills) can contain prompt injection
• Credentials in plaintext config files have no access scoping or rotation

This skill applies three frameworks:

1. Microsoft ZT4AI — Verify explicitly, least privilege, assume breach
2. "Caging the Agents" (arXiv:2603.17419) — Four-layer defense: workload isolation, credential proxy, network egress, prompt integrity
3. OWASP Agentic AI Top 10 — Trust boundary violations, privilege escalation, resource exhaustion

Audit Process