Agentic AI Security Career Roadmap
Why this roadmap exists
Most “AI security” roadmaps stop at a job list and a reading list. They tell you which titles are hiring and which frameworks to memorize, but they leave out the parts that actually get people hired: a current picture of the threat landscape, the governance frameworks that turn findings into decisions, the community and credential signals that make a stranger trust your work, a concrete path from portfolio to offer, and a learning loop that keeps you relevant after the first 90 days.
This site fills those gaps. It is organized as one continuous workflow — learn → build → break → document → get hired → keep growing — split into six sections.
- Career Paths — the five core Agentic AI Security roles, three fast-growing adjacent roles, and how to choose based on where you are coming from.
- Skill Tree — five layers, from security fundamentals and engineering craft up through AI tooling and the soft skills that decide whether your findings get acted on.
- Threat Landscape — what actually breaks: prompt injection, memory poisoning, tool misuse, identity spoofing, and the incident patterns behind them.
- Governance & Compliance — NIST AI RMF, the EU AI Act, and ISO/IEC 42001, plus how to collect compliance evidence automatically instead of by hand.
- 90-Day Plan — a month-by-month build schedule with two agents you ship, one red team lab you run against yourself, and the artifacts you publish.
- Career Launch & Growth — certifications worth the money, converting a portfolio into interviews, and the weekly/monthly/quarterly/annual loop that prevents the day-91 plateau.
Recommended order: ③ → ① → ② → ⑤ → ④ → ⑥
Start with the threat landscape, not the skill list. The fastest way to know which skills you actually need is to first see how agentic systems fail — an agent that reads an attacker-controlled web page, calls a tool with borrowed credentials, and writes the result into its own long-term memory is a very different security problem from a misclassified image.
Once the failure modes are concrete, Career Paths tells you which role owns which failure, Skill Tree tells you what to study to own it, and the 90-Day Plan forces you to build something that fails in those exact ways so you can document it. Governance comes late on purpose: frameworks are far easier to internalize once you have a real system to map them onto, and they are what turns a demo into an artifact an enterprise reviewer takes seriously.
Where to start, by background
The 12-week plan at a glance
| Weeks | Focus |
|---|---|
| 1–4 | Python, Git, Docker, and an AI coding assistant. Ship the Log Triage Agent. Self-assess against at least five OWASP LLM Top 10 entries. |
| 5–8 | Orchestration, APIs, cloud security, and MCP. Ship the Multi-Agent SOC and expose its tools through an MCP server. |
| 9–10 | Red team your own agents: prompt injection, memory poisoning, tool misuse. Capture everything as audit logs and write it up. |
| 11–12 | Publish 3–5 projects with a STRIDE threat model, a trust boundary diagram, an OpenTelemetry audit trail sample, and a “how this agent fails” section. |
The detailed week-by-week version lives in the 90-Day Plan; the runnable versions of these projects live in Labs.