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

Best practiceRunning agents

Require a human for consequential actions

OWASP ASI02OWASP ASI05OWASP ASI09

Payments, deletions, sending messages, changing permissions, pushing code and modifying infrastructure are one-way doors. Approval prompts are the last control between a hijacked agent and real damage, and every bypass of them has become a CVE.

What to do

  1. Classify tools as read, write or consequential; consequential ones always prompt, with the exact arguments shown.
  2. Never run agents in auto-approve mode (YOLO, --dangerously-skip-permissions, headless trust) on machines with credentials or production access.
  3. Keep allow-lists narrow and exact: a permitted command with flags (find -exec, npx, git push --receive-pack) is a permitted shell.
  4. Update agent tools promptly; approval bypasses are fixed often and quietly.

Sources

Check your own AI plan against the practices: Plan check