25-Point AI Security Checklist
The same checklist our CREST-certified team uses to assess enterprise AI systems. Based on the OWASP Top 10 for LLMs, NIST AI RMF, and lessons from the McKinsey Lilli breach.
Covers the vulnerabilities that led to the McKinsey Lilli breach (46.5M messages exposed)
Includes AI coding tool security (Claude Code, Copilot, Cursor)
Maps to EU AI Act compliance requirements (deadline: August 2, 2026)
73% of AI deployments fail this checklist
According to OWASP, prompt injection vulnerabilities appear in over 73% of production AI systems assessed during security audits.
Get the Checklist
Free instant download — no spam, ever
What's Inside the Checklist
AI API Security (5 checks)
All AI API endpoints require authentication (no unauthenticated endpoints)
Role-based or attribute-based access control (RBAC/ABAC) enforced on every endpoint
Rate limiting and throttling active on all AI-facing APIs
API documentation is not publicly accessible
Input validation and output sanitisation applied to all AI endpoints
Prompt & Model Security (5 checks)
System prompts stored separately from user-accessible databases
System prompt integrity monitoring in place
Prompt injection defences tested (direct, indirect, multi-turn)
Output filtering prevents PII leakage and harmful content generation
Model behaviour monitoring detects anomalous responses
Data & Privacy (5 checks)
AI input/output data flows are documented and mapped
Data minimisation principles applied — AI only accesses data it needs
Prompt history and chat logs encrypted at rest and in transit
PII handling compliant with GDPR/PDPA/HIPAA as applicable
Third-party AI service data processing agreements reviewed
RAG & Training Pipeline (5 checks)
RAG document access controls enforce user-level permissions
Cross-tenant data isolation validated in multi-user RAG systems
Training data provenance verified — no tainted data sources
Vector store access restricted and monitored
External data ingestion validated for injection payloads
AI Tool & Compliance Governance (5 checks)
Inventory of all AI tools in use across the organisation (including Claude Code, Copilot, etc.)
AI tool configurations in repositories reviewed as part of code review process
API keys and credentials for AI services rotated regularly and monitored
EU AI Act risk classification completed for all AI systems
Adversarial testing (red teaming) scheduled before August 2, 2026 compliance deadline