SaaS Agent Risk & Stress-Test Skill

Overview

AI risk registers (handled by saas-ai-risk-and-stress-test) cover cost spike, model deprecation, hallucination event, data-rights conflict, vendor concentration, regulatory exposure. Agent risk must add two categories that do not exist in AI-feature plans:

1. Autonomy incidents — the agent takes an action the customer / regulator considers wrong, beyond authority, or harmful
2. Irreversibility incidents — the agent takes an action that cannot be undone (financial transfer; record committed to ledger; communication sent; legal document filed; permit issued; medication ordered)

Plus agent-specific failure modes: multi-agent collusion / loop, tool-vendor outage breaking the action layer, prompt-injection escalation that bypasses safety, action-authentication bypass, AI Safety Lead flight, foundation-model deprecation that breaks specific tool-call signatures.

Generic risk skills miss all of these.

Use When