AI Security & Red Teaming — Inference Phase

🔐

Inference Specialization

AI Security & Red Teaming

AI Security engineers protect systems from adversarial attacks, prompt injection, data poisoning, and model theft. Gartner forecasts $2.5 trillion in AI cybersecurity spending by 2026. A genuinely new discipline with almost no trained specialists available.

Days

Projects

$185K

Median

$2.5T

Market by 2026

$130K–$250K

Salary Range

Week 1: Adversarial ML and Attack Taxonomy · Days 1–8  

Understand the complete threat surface of ML systems: adversarial examples, data poisoning, model extraction, and LLM-specific attacks.

D1–3

Adversarial ML: evasion, poisoning, and model extraction

FGSM, PGD, C&W adversarial examples. Data poisoning attacks, backdoor insertion, model stealing via queries

↗ IBM Adversarial Robustness Toolbox documentation (free)

D4–5

LLM-specific attacks: prompt injection, jailbreaks, denial of service

Direct and indirect prompt injection, jailbreak taxonomy, token flooding, context window manipulation techniques

↗ Red Teaming LLM Applications — DeepLearning.AI (free)

D6–7

AI threat modelling: MITRE ATLAS and STRIDE for ML

Applying STRIDE and DREAD to ML pipelines. MITRE ATLAS attack taxonomy, attack trees for LLM applications

↗ MITRE ATLAS AI attack framework (free)

🏗 Project — Day D8: Adversarial Attack Suite on a Production AI System 

· FGSM and PGD attacks on a production image classifier

· 5 prompt injection attacks on a RAG chatbot with success rates

· MITRE ATLAS threat model document for a real AI system

· Attack success rate report with prioritised defence recommendations

Week 2: Defences, Red Teaming Processes, and System Hardening · Days 9–24  

Build effective defences: adversarial training, input validation layers, guardrails, and a systematic red-team methodology.

D9–12

Adversarial training and certified robustness techniques

Adversarial data augmentation during training, randomised smoothing, Lipschitz-constrained neural networks

↗ ART adversarial training tutorials (free)

D13–16

LLM guardrails and content safety architecture

Llama Guard, NeMo Guardrails, OpenAI Moderation API. Multi-layer defence-in-depth architecture patterns

↗ Meta Llama Guard GitHub and documentation (free)

D17–20

Systematic red teaming: methodology, tooling, and reporting

Red team playbook development, Garak automated vulnerability scanning, PyRIT, structured vulnerability disclosure

↗ Garak LLM vulnerability scanner (free)

D21–23

Supply chain security: model poisoning and dataset integrity

Scanning models on HuggingFace Hub, dataset provenance auditing, checksum verification, secure registries

↗ HuggingFace Hub Security documentation (free)

🏗 Project — Day D24: Security Hardening Report for an LLM Application 

· Full Garak scan with 50 or more attack probe categories

· MITRE ATLAS threat model diagram with attack paths

· Implemented guardrails layer with before and after metrics

· Responsible disclosure write-up template for AI vulnerabilities

Week 3: AI Governance, Compliance, and Advanced Topics · Days 25–40  

AI governance frameworks, regulatory compliance requirements, privacy attacks and defences, and building a complete AI security programme.

D25–28

AI governance: NIST AI RMF and EU AI Act

NIST AI Risk Management Framework tiers, EU AI Act risk classifications, ISO 42001, corporate AI policies

↗ NIST AI Risk Management Framework (free public document)

D29–32

Privacy attacks: model inversion and membership inference

Extracting training data from deployed models. Differentially private training as a systematic defence

↗ OpenDP differential privacy platform (free)

D33–36

Secure ML pipelines: supply chain, Zero Trust, and secrets management

Zero Trust Architecture for ML infrastructure, secrets management, signing model artefacts, SLSA compliance

↗ SLSA supply chain security framework (free)

D37–39

Bug bounty programmes and responsible disclosure for AI

AI vulnerability research methodology, CVE process for ML issues, building an internal AI bug bounty programme

↗ MITRE ATLAS case studies and reports (free)

🏗 Project — Day D40: AI Security Audit and Hardening Programme — Full Capstone 

· Complete security audit of a real AI application or previous project

· Garak automated plus manual red team findings report

· NIST AI RMF compliance checklist with gap analysis

· Implemented mitigations with before and after vulnerability metrics

· Security runbook for the development and operations team