Zero Trust AI Agents Skill

An Agent Skill for assessing, threat modeling, and hardening autonomous AI agent deployments with Zero Trust and Least Agency principles.

This skill is designed for tools that support SKILL.md-based skills, including Claude Code and Codex-style skill loaders. Claude's public docs describe skills as folders containing a required SKILL.md plus optional scripts and references, discoverable from personal and project skill directories. Codex skills use the same core shape, with optional agents/openai.yaml metadata for UI surfaces.

What It Does

• Reviews agent architectures, MCP/tool access, credentials, memory/RAG, observability, governance, and defensive operations.
• Produces Zero Trust maturity assessments across Foundation, Enterprise, and Advanced tiers.
• Generates implementation roadmaps, blast-radius analyses, control matrices, and evidence requests.
• Includes a no-dependency Python script for repeatable inventory-based assessments.

Install

For Claude Code personal skills:

mkdir -p ~/.claude/skills
cp -R zero-trust-ai-agents ~/.claude/skills/

For Claude Code project skills:

mkdir -p .claude/skills
cp -R zero-trust-ai-agents .claude/skills/

For Codex personal skills:

mkdir -p ~/.codex/skills
cp -R zero-trust-ai-agents ~/.codex/skills/

Many tools reload skills automatically when files change. If your agent does not see the skill, restart the session so metadata is reloaded.

For Claude.ai custom skills, zip the contents of this directory so SKILL.md is at the archive root, then upload the zip through the Skills UI.

Example Prompts

Use $zero-trust-ai-agents to review this MCP-based support agent for production readiness.

Use $zero-trust-ai-agents to create a Foundation-to-Enterprise roadmap for our research agent.

Use $zero-trust-ai-agents to threat model prompt injection, tool misuse, and memory poisoning for this agent architecture.

Inventory Assessment

Run the bundled script:

python scripts/zt_agent_assessment.py examples/agent-inventory.example.json --out assessment.md

The inventory schema is intentionally simple JSON. Use booleans for controls, or values like "partial" / "planned" when evidence is incomplete.

Repository Layout

zero-trust-ai-agents/
├── SKILL.md
├── agents/openai.yaml
├── examples/agent-inventory.example.json
├── references/control-catalog.md
├── references/implementation-workflow.md
├── references/output-templates.md
├── references/threat-patterns.md
├── scripts/zt_agent_assessment.py
└── tests/test_skill_package.py

Test

python -m unittest discover -s tests

Notes

This project paraphrases and operationalizes concepts from Anthropic's "Zero Trust for AI Agents" ebook. It does not copy the ebook text and does not provide legal, compliance, or security assurance.

References

• Claude Code Agent Skills docs: https://docs.claude.com/en/docs/claude-code/skills
• Claude Skills overview: https://claude.com/skills