AI Security Resource
AI Security Architecture Review Checklist
Use this checklist to prepare for an independent review of an enterprise AI system, RAG environment, AI agent, MCP integration, or AI-enabled delivery pipeline.
What to prepare before a review
The strongest reviews start with enough context to understand how the AI system is built, what it can access, how changes are shipped, and which risks the business is trying to reduce.
Evidence handling note
Do not send confidential diagrams, credentials, regulated data, production logs, or exploit details through email. Start with a high-level description, then arrange an approved secure channel if sensitive evidence is needed.
Checklist areas
Use these categories to identify what is ready, what is missing, and which stakeholders should be involved.
Architecture and data flow
- System context diagram
- Data-flow diagram
- Trust boundaries
- External integrations
- Deployment environments
Identity and authorization
- User and service identities
- Agent permissions
- Tool access rules
- Approval workflows
- Privileged action controls
RAG and knowledge sources
- Source repositories
- Document access model
- Chunk metadata
- Retrieval filtering
- Data leakage controls
Prompt and model controls
- System instructions
- Prompt change process
- Guardrails and policies
- Evaluation results
- Fallback behavior
MCP and tool integrations
- MCP server inventory
- Tool scope
- Credential handling
- Logging and audit trails
- Network exposure
Attack simulation readiness
- Written authorization boundaries
- Rules of engagement
- Prompt injection scenarios
- Tool abuse test cases
- RAG exposure test cases
AI CI/CD and operations
- Pipeline permissions
- Secrets and artifacts
- Approval gates
- Monitoring coverage
- Incident response ownership
Questions to answer internally
- What business decision or workflow does the AI system influence?
- Which data can the system retrieve, transform, summarize, or expose?
- What actions can agents or tools perform without human approval?
- How are prompts, policies, evaluations, and model settings changed?
- Which logs would help investigate prompt injection, tool abuse, or data exposure?
- Who owns risk acceptance and remediation decisions?
Ready to discuss scope?
Send a non-confidential overview to contact@assured-ai-security.com with the system type, project stage, desired timeline, and the outcome you want from a review.
Email General Consulting Inquiry