Test tool-using agents before launch

Test tool boundaries, permission checks, and confirmation flows.

The risky part is not what the agent says. It is what it can do next.

What can go wrong

• Tools run without confirmation.
• Users escalate privilege through claimed roles.
• Indirect prompts trigger refunds or data exports.

Why happy-path testing misses this

• Tool success paths are tested. Abuse paths often are not.

What Agnostics tests

• Unsafe tool actions
• Permission abuse
• Boundary bypass

Useful finding example

Permission abuse finding: export attempted after fake admin claim.

Release Gate

Fix or Blocked for launch-critical workflows with open action findings.