Skip to content
Pedro Mora
Go back

AI Security Competency Matrix

active research Mar 2026 – present

Most AI security content is opinion. This is evidence.

The matrix is a structured learning system across 9 domains — from IAM for AI systems to LLM threat mitigations, agent architecture, platform security, and governance. Each competency has three levels: Beginner, Intermediate, Expert. Progress is tracked with lab work and published writeups as evidence.

152 competencies. 9 domains. All of it documented here.


How It Works

Each competency moves through four states:

BadgeStateMeaning
🔲UnassessedHaven’t looked at it yet
📖LearningActively studying, notes in progress
🧪TestedBuilt a lab or ran a test
PublishedWritten up and live on this site

Published posts are the evidence layer — not self-assessment, not a certificate. If the writeup exists, the competency is covered.


Domain Progress

D1 — Identity & Access Management for AI Systems

Beginner: 📖 in progress

IAM is the foundation. Before you can reason about AI system security, you need fluency with identity lifecycle, least privilege, and access control patterns — in practice, not in theory. The Okta cert labs are the hands-on evidence layer for this domain.

Competencies in progress: B·C2 (OAuth 2.0 / OIDC relevance for AI agents) · B·C4 (least privilege with AI agent example) · B·C5 (audit logging for AI agent actions) · B·C6 (identify misconfigured AI access controls)

Evidence: Okta Cert Prep labs →


D2 — AI/LLM Security Threats & Mitigations

Beginner: 📖 in progress

The OWASP LLM Top 10 is the threat vocabulary. The HTB OSAI Prep series builds it from the bottom up — classical offensive techniques first, then the structural equivalent in AI deployments. Prompt injection (LLM01), sensitive information disclosure (LLM06), insecure plugin design (LLM07), and excessive agency (LLM08) all have live evidence.

Competencies covered: B·C1 (explain prompt injection with a simple example) · B·C3 (explain PII leakage from AI systems)

Evidence: OSAI Prep writeups →


D3 — AI Agent Architecture & Control Frameworks

Beginner: 🔲 not started

Agentic loops, tool use, human-in-the-loop controls. Next domain after D1 and D2 beginner levels are complete.


D4 — LLM Technical Internals

Beginner: 🔲 not started

Transformer architecture, context windows, tokenization, RAG. The technical substrate that makes the threats in D2 possible.


D5 — AI Platform Security

Beginner: 🔲 not started

Azure OpenAI, AWS Bedrock, Anthropic API. Platform-specific deployment and access control patterns.


D6 — AI Governance & Compliance

Beginner: 🔲 not started

EU AI Act, model risk management, audit logging requirements. The regulatory layer.


D7 — Statistical Methods & ML Theory

Beginner: 🔲 awareness target only

This domain has a ceiling. The goal is functional literacy for conversations with ML engineers — not a deep dive.


D8 — AI Supply Chain Security

Beginner: 🔲 not started

Poisoned models, model provenance, third-party AI library risk.


D9 — Thought Leadership & Communication

Beginner: 📖 in progress

Every published post on this site is evidence for D9 — explaining AI security concepts clearly, in writing, to a mixed audience. B·C2 (write a clear AI security risk summary for a non-security audience) is covered every time a writeup goes live.

Evidence: All posts →


What’s Next

Working through D1 and D2 Beginner levels in parallel — Okta cert labs for D1, OSAI Prep HTB writeups for D2. First full domain-level post (D2 Beginner roundup) when the HTB series covers the remaining OWASP LLM Top 10 entries.