19 weeks · 111 study days · 555 hours
The whole roadmap
Six study days per full week, one rest day after each, and three days in week 19. Every day is five hours: concepts, build, security lab, evidence, Socratic review. If a day is missed, move it forward — never compress two days into one.
Week 1
Application foundations, model trust, and first local assistantS0S1S3S70/6
Week 2
Tools, prompt injection, and authorization boundariesS2S30/6
Week 3
Streaming, reliability, cost controls, and memory foundationsS40/6
Week 4
Embeddings, model artifact integrity, and retrieval foundationsS1S2S40/6
Week 5
RAG evaluation, privacy, and training-data securityS3S4S50/6
Week 6
Agent foundations, MCP, and protocol securityS3S60/6
Week 7
Agent state, identity, delegation, and approvalS0S60/6
Week 8
Background jobs, SSRF, economic abuse, and privacyS00/6
Week 9
Observability, AI-SDLC, secure release, and red-team automationS7S8S90/6
- D49Add container scanning and model-artifact inventory to the…D12M1M8
- D50Build a compact security regression suite for prompt injectionD8M8S8
- D51PyRIT / Garak Scheduled FuzzingD8M8S8
- D52Implement output validation with Pydantic/JSON SchemaD6S7
- D53Study grammar-constrained generation and structured decodingD6S7
- D54Model Extraction Threat ModelS9
Week 10
Output security, model extraction, IP protection, and governanceS110/6
Week 11
Browser/computer-use agent security and end-to-end attack chainsS100/6
Week 12
Mechanistic interpretability, robustness, and research literacyS10S120/6
Week 13
Security teardown, CVEs, responsible disclosure, and architecture reviewS0S1S2S3S50/6
Week 14
Project 1 hardening: assistant as a secure product0/6
Week 15
Project 3 hardening: agent/workflow as a secure product0/6
- D85Agent State Machine & Sandboxed Code DiagnosisD2D3M5S3
- D86Review OAuth2/delegated scopesD7S6
- D87Review memory/checkpoint provenanceD10D11M6S11
- D88Browser Sandbox & Pixel Auditor AuditD2D3D7D8D9D11M9M10
- D89Create a unified attack taxonomy across projectsD1D2D3D4D5D6D7D8D9D10D11D12S1S8
- D90Run direct, indirect, multi-turn, encoded, split-payload, and…D1D8S2S8
Week 16
Cross-project red teaming and security evaluationS0S3S4S6S80/6
- D91Authorization & Isolation Red TeamingD2D3D4D5D7D10D11D12M1S0S4
- D92Run availability and economic abuseD1D8M5S9
- D93Prioritize findings by impact and exploitabilityD8D11D12S8S11
- D94Design incident severity levels for data exposureM10S11
- D95Drill 1: Compromised Credential / Delegated TokenD7D8D11D12M1S6
- D96Simulate a poisoned document or memory entryD5D10M6S4
Week 17
Reliability, incident response, and operational securityS110/6
- D97Simulate agent runaway or unsafe browser actionD2D7D9D11M9S3
- D98Incident Playbooks & Tabletop ExerciseS11
- D99Review EU AI Act concepts relevant to the intended use caseS11
- D100Build a third-party inventoryD12M1S11
- D101Enterprise Audit & Retention PolicyD8D11D12M8S11
- D102Complete a governance/risk review with ownersS11
Week 18
Advanced labs and targeted remediation0/6
- D103Revisit the weakest model supply-chain controlD12M1
- D104Revisit training-data poisoning/backdoor experimentD5D10M2S4
- D105Multimodal Conflict HardeningD2D5D12M3M4
- D106Revisit advanced privacy and deletionD4D10M7S5
- D107Revisit browser/computer-use securityD9M9
- D108Project 1 Clean-Room RunD1D4D5D6D8D10D11D12
Week 19
Final integration, portfolio evidence, and competency assessment0/3
111 days · every source block preserved verbatim