Skip to the content
AI and Cybersecurity Learning and Practice Center Test automation Open the Learning Center

Best practiceGovernance and people

Keep an inventory of models, agents, tools and their permissions

OWASP ASI04OWASP ASI10

You cannot secure what you do not know exists. Shadow agents, personal MCP installs and forgotten evaluation endpoints are how attackers found DeepSeek's open database and thousands of exposed Ray, Ollama and n8n instances.

What to do

  1. Record every model, agent, MCP server, plugin and connector with its owner, version, data access and credentials, as an AI bill of materials.
  2. Discover what is actually running, shadow AI included: scan your networks and clouds for AI service ports and agent processes.
  3. Rate each system's risk the way the EU AI Act sorts AI (banned, high-risk, must be transparent, minimal), and let the rating set how much review it gets.
  4. Tie each entry to a risk owner and a review date; retire what no one claims.

Sources

Check your own AI plan against the practices: Plan check