What problem does it solve? Standard capability benchmarks like MMLU never test what a model can be made to do, leaving prompt injection, jailbreaks, system-prompt extraction, and data disclosure undetected until users find them. This Skill structures an authorized, bounded red-team engagement and turns every confirmed vulnerability into a durable regression eval case. ## Core Features & Use Cases - Attack-surface enumeration and battery assembly: Covers bias elicitation, sensitive-information disclosure, service disruption, and hallucination induction, with injection/jailbreak/probe techniques drawn from OWASP LLM Top 10, AVID, and the AI Incident Database. - Automated scanning: Runs Giskard LLM scans as scheduled Kubernetes Jobs against staging or live endpoints, with rate, spend, and kill-switch controls. - Three-axis severity classification: Classifies each finding by influence (causative vs exploratory), security goal (integrity, availability, privacy), and specificity, then maps it to a blue-team defense such as a guardian LLM relevance scorer. - Use Case: Before promoting a fine-tuned model, run the full battery against the staging checkpoint, classify findings, and hand the critical-findings verdict to the release gate so a model with an open critical finding cannot ship. ## Quick Start Use the genai-red-team skill to scope and run an authorized red-team engagement against my staging model endpoint and produce the engagement report.
