What problem does it solve? LLM and ML systems face threats like prompt injection, jailbreaks, model inversion, and data poisoning that traditional application security tools do not detect. This Skill provides a structured methodology and a scanner tool to assess AI/ML systems against these threats and map findings to MITRE ATLAS techniques. ## Core Features & Use Cases - Prompt Injection & Jailbreak Detection: Scan built-in seed prompts or custom JSON test files against regex-based injection signatures covering role overrides, indirect injection, persona jailbreaks, system prompt extraction, and tool abuse. - Risk Scoring & ATLAS Mapping: Compute injection scores, model inversion risk by access level, data poisoning risk by fine-tuning scope, and map every finding to MITRE ATLAS technique IDs. - CI/CD Security Gate: Use exit codes (0/1/2) to block deployments when critical AI security findings are detected. - Use Case: Before deploying a customer-facing LLM feature, run the scanner against your domain-specific adversarial prompts, review the overall risk level, and implement the recommended guardrails such as input filters and tool approval gates. ## Quick Start Run the AI threat scanner against the built-in seed prompts for a black-box LLM target and review the JSON risk report.
