AI & Automation Engineering
The tooling layer that scales the rest.
Security work is dominated by tasks that are high-volume, judgment-heavy, and poorly served by deterministic automation: triage, enrichment, summarization, evidence collection, first-pass code review. That’s a good fit for language models, and a catastrophic fit for language models used carelessly — because in security, a confident wrong answer is worse than no answer.
So this pillar has two halves that have to be read together: building AI systems that work, and building them so their failures are survivable.
Why it matters
Section titled “Why it matters”The constraint on most security programs is analyst hours, not knowledge. Automation is the only lever that changes that ratio. But it changes it only if the output is verifiable — which is why evaluation and grounding get as much space here as retrieval and orchestration.
How this connects
Section titled “How this connects”- → IAM: agent identity is the same problem as §2.4 Zero Knowledge Trust. An agent with tools is a principal with permissions, and it needs a credential lifecycle, scoping, and revocation.
- → Detection Engineering: agent-assisted triage and SOAR enrichment are the production use case — with human approval gates on anything destructive.
- → GRC: automated evidence pipelines are the highest-ROI application in security, and the EU AI Act makes the models themselves a compliance surface.
- → Threat Intel: summarization pipelines scale collection, and hallucinated intel is the failure mode that discredits the whole program.
- → AppSec: scanning augmentation and code review assistance, plus prompt injection as a genuinely new vulnerability class.
- → Penetration Testing & Red Teaming: LLM applications are a target class with their own methodology.
- → Cloud & Infrastructure Security: drift detection and closed-loop remediation are a natural automation target, and model-serving infrastructure is a workload with an unusual blast radius.
- → Incident Response & Digital Forensics: triage and enrichment are the obvious automation targets, with human approval gates on anything destructive.
- → Data Security & Privacy Engineering: training-data provenance, PII in prompts and embeddings, and deletion once data is baked into weights. A vector store is a data store nobody classified.
What’s here
Section titled “What’s here”| Subsection | Focus |
|---|---|
| Retrieval-Augmented Generation | Naive RAG failure modes, query rewriting/HyDE/hybrid/reranking, agentic and GraphRAG, contextual retrieval, long-context vs. retrieval, chunking and evaluation |
| Vector Databases & Embeddings | What embeddings encode, similarity metrics and index types (HNSW, IVF, PQ), the DB landscape, hybrid search and fusion, metadata filtering, cost/latency/recall |
| Verification & Evaluation | Grounding and citation enforcement, hallucination detection, LLM-as-judge pitfalls, golden datasets, guardrails and schema validation, human-in-the-loop |
| Agent Orchestration | Single vs. multi-agent, planner/executor and supervisor and reflection patterns, LangGraph/CrewAI/AutoGen/MCP, memory, loop termination, cost control |
| Security Applications | Triage and alert enrichment, report drafting, code review augmentation, compliance evidence, intel summarization |
| Securing AI Systems | Direct and indirect prompt injection, OWASP Top 10 for LLM Applications, tool-use authorization and blast radius, model supply chain |