Look beyond the model
Data, retrieval, identity, tools, application logic, people, and operations determine the real risk.
Learn / AI security
AI Security Field Guide
Learn to test, defend, and govern AI systems without crossing the line.
A practical, beginner-friendly learning path for AI security: foundations, safe labs, machine learning, LLM threats, red teaming, detection, cloud controls, and governance.
Modules14
From foundations to incident response.
Learning paths3
Foundation, testing, and defense.
Practice modelSafe by default
Local, synthetic, or explicitly authorized.
UpdatedJuly 31, 2026
Current frameworks and links checked.
How this is different
A security-practitioner guide, not an exploit cookbook
The source PDF supplied a useful beginner topic map. This adaptation replaces unsafe live-target examples, outdated tool snippets, and unqualified claims with system thinking, controlled labs, current terminology, and evidence-led practice.
Verify every result
Models generate hypotheses. Reproduction, logs, tests, and source evidence turn them into findings.
Test consequences
Prioritize data exposure, authorization failure, and side effects over clever prompt demonstrations.
Design for recovery
Every important control needs an owner, signal, stop path, retest, and rollback.
Choose your route
Three connected learning paths
Start from zero
Build the vocabulary, lab discipline, Python confidence, and model-evaluation habits needed for every later module.
5 modulesTest AI safely
Threat-model an AI application, test prompt boundaries, review code, and run repeatable evaluations inside a documented scope.
5 modulesBuild and operate securely
Design detection, constrain agents, secure the model supply chain, harden cloud deployment, and prepare response evidence.
4 modulesFoundation
5 modules
AI security starts with the system, not the model
Learn the difference between using AI for security, securing an AI application, and protecting the model lifecycle.
Build a lab that cannot surprise production
Create an isolated workspace, a written scope, synthetic data, and a stop procedure before running security experiments.
Use Python to inspect security data, not to hide reasoning
Learn the small set of Python and data-handling patterns needed to explore logs, build features, and preserve an audit trail.
Measure the security decision, not just model accuracy
Build a practical mental model of supervised learning, anomaly detection, thresholds, drift, and the cost of false decisions.
Understand the application wrapped around the LLM
Follow a request through instructions, retrieval, tools, memory, output parsing, and the final business action.
Test safely
5 modules
Threat-model AI across data, model, application, and action
Turn a system map into abuse cases, control hypotheses, and a risk-based test plan.
Test prompt injection as a control failure
Move beyond jailbreak tricks and test whether untrusted content can cross an authorization or action boundary.
Use AI to accelerate secure code review without inventing findings
Combine model-assisted review with static analysis, tests, dependency evidence, and developer context.
Understand evasion, poisoning, privacy, and model extraction
Use the NIST taxonomy to reason about adversarial machine learning across training, deployment, and inference.
Make AI red teaming repeatable, scoped, and decision-relevant
Combine human creativity, automated probes, deterministic checks, and regression suites without mistaking volume for coverage.
Defend and operate
4 modules
Design detection around behavior, evidence, and analyst capacity
Build anomaly and AI-system monitoring that produces actionable investigation paths instead of unexplained scores.
Constrain AI agents with identity, policy, and small tools
Design an agent runtime where untrusted instructions cannot silently become broad authority.
Secure the AI supply chain and cloud runtime
Protect models, datasets, notebooks, registries, pipelines, endpoints, secrets, and observability across the deployment lifecycle.
Prepare evidence, response, disclosure, and governance before failure
Connect technical controls to incident handling, responsible disclosure, risk ownership, and the EU operating context.
Non-negotiable
The field-guide safety contract
Every practical exercise in this section assumes these boundaries. If a lab cannot meet them, stop and redesign the lab.
- Use systems you own, intentionally vulnerable training applications, synthetic data, or targets covered by explicit written authorization.
- Define scope, permitted techniques, stop conditions, data handling, and an emergency contact before any assessment.
- Treat model output as a hypothesis. Verify findings with deterministic tools, logs, source code, and human review.
- Keep credentials, production data, customer information, and real personal data out of prompts and lab datasets.
- Prefer evidence that improves a control or a fix over demonstrations that merely prove an attack is possible.
Source and adaptation
What came from the document
The user-provided 2026 PDF supplied the initial topic map. This field guide is an original rewrite with a defensive-first structure, current sources, safer labs, corrected terminology, and enterprise operating context.
Visit Codelivly