ai-incident-response

What problem does it solve?

This Skill helps teams rapidly classify, mitigate, and learn from AI-specific incidents such as model degradation, hallucination spikes, bias detections, safety bypasses, prompt injection, and data poisoning, reducing user harm and legal exposure.

Core Features & Use Cases

• Context confirmation & triage: Guided questions to confirm affected feature, blast radius, and whether the incident is active or preparatory.
• Incident taxonomy & severity: Clear classification of incident types and severity levels (P0–P3) with SLAs for response.
• Response runbooks: Step-by-step mitigations per incident type (rollback, input filters, grounding, human review), rollback procedures, and communication templates.
• Post-incident analysis & monitoring: 5-whys adapted for AI, post-mortem deliverables, and recommended real-time and near-real-time metrics for early warning.
• Use Cases: Responding to an active hallucination spike in production, preparing an AI safety incident playbook for a new feature, or drafting a P0 bias communication and remediation plan.

Quick Start

Respond to an active hallucination spike in the recommendations feature affecting 10% of users by classifying severity, proposing immediate mitigations (rollback or grounding), and outlining a post-incident action plan.