ai-security-vuln-scan

name	ai-security-vuln-scan
description	Deterministic workflow for AI security vulnerability scanning across code repos, dependency manifests, container images, infrastructure-as-code, CI/CD configs, and LLM applications. Use when asked to run AI/LLM security reviews, triage vulnerabilities, map findings to OWASP LLM Top 10, assign remediation SLAs, and produce evidence-ready audit artifacts.

AI Security Vulnerability Scan

Goal

Produce repeatable, evidence-ready vulnerability assessments for AI systems.

Required inputs

• target_type: repo, container, iac, ci-config, llm-app, or mixed
• target_path_or_ref: path, image tag, or config location

Optional inputs

• scan_depth: quick, standard, deep (default: standard)
• policy_profile: default owasp-default
• business_context: crown-jewel designation, regulated data, environment

Workflow

1. Scope targets.
   • Inventory all in-scope assets.
   • Mark production vs non-production.
   • Capture data classes (PII/PHI/PCI/IP) when known.
2. Run intake.
   • Fill templates/intake-questionnaire.md.
   • Ask only blocking follow-ups if required inputs missing.
3. Execute layered checks.
   • For code/dependencies: scan vulnerable packages, secrets exposure, unsafe deserialization, injection paths.
   • For containers: scan base image CVEs, root user usage, exposed ports, package drift.
   • For IaC: scan overly broad IAM, public exposure, missing encryption/logging.
   • For CI/CD: scan token scope, unpinned actions, artifact trust boundaries.
   • For LLM app: test prompt injection, tool overreach, insecure output handling, sensitive data leakage, excessive agency.
4. Record findings deterministically.
   • Use one templates/finding-record.md entry per finding.
   • Populate severity and exploitability from references/severity-model.md.
   • Map every finding to references/owasp-llm-mapping.md.
5. Triage and assign remediation.
   • Sort findings by risk score, then exploitability.
   • Assign owner and SLA using templates/remediation-sla-matrix.md.
   • Flag false positives only with written evidence and expiration.
6. Build evidence pack.
   • Complete templates/evidence-pack-index.md.
   • Attach logs, scan config, commit/image refs, and validation proof.
7. Set release gate status.
   • Mark release_blocker=true if unresolved critical finding exists.
   • Mark release_blocker=true if unresolved high finding has a clear exploit path to sensitive data or privileged tools.
8. Define regression checks.
   • Add at least one verification check per closed high/critical finding.

Output contract

Return sections in exact order:

1. Scan Context
2. Target Inventory
3. Prioritized Findings
4. OWASP Mapping Summary
5. Owners and SLA
6. Evidence Pack Index
7. Residual Risk and Release Gate
8. Regression Validation Plan

Required fields per finding

• id
• title
• asset
• severity
• exploitability
• risk_score
• owasp_category
• proof
• owner
• sla_due
• status

Guardrails

• Keep output free of secrets and tokens.
• Prefer deterministic evidence over narrative claims.
• State uncertainty explicitly when data is missing.
• Recommend human approval gate for high-impact production changes.