AI Security Engineering

Course 01

Securing LLM applications

The new attack surface an LLM feature opens, and how to build one that does not leak its system prompt, its data or its keys.

4 lessons · ~7 h

Tools Claude API, an LLM app framework

1. 01The LLM attack surfaceWhere an LLM feature widens the surface; trust the model gives you and does not.90 min
2. 02Input and output handlingTreating model input and output as untrusted; the boundaries to enforce.120 min
3. 03Context isolationKeeping tenant data, system prompts and secrets out of reach of the prompt.120 min
4. 04Keys and cost controlProtecting the API key and bounding spend against abuse.90 min

Course 02

Prompt injection and tool sandboxing

The defining LLM vulnerability. Direct and indirect injection, and sandboxing the tools a model can call so a hijacked prompt cannot become a hijacked system.

3 lessons · ~6 h

Tools Tool isolation, capability scoping

1. 01Direct and indirect injectionInjection from the user and from retrieved content; why the model cannot tell them apart.120 min
2. 02Tool sandboxingEvery tool scoped to least privilege; a hijacked prompt confined to what the tool may do.120 min
3. 03Defenses that holdIsolation, allow-lists and human-in-the-loop where the action is dangerous.90 min

Course 03

Agent and MCP security

Autonomous agents that act on the world, and the Model Context Protocol that connects them to tools and data, secured so an agent cannot exceed its authority.

3 lessons · ~6 h

Tools MCP, agent frameworks, authorization

1. 01Agents that actThe risk when a model does things, not just says things; blast radius by design.90 min
2. 02MCP securityAuthorizing and auditing MCP tool calls; the trust model of the protocol.120 min
3. 03Bounding authorityScopes, approvals and audit so a manipulated agent stays inside its lane.120 min

Course 04

AI for security operations

Turn the technology into a defender. LLM-assisted triage, detection and investigation, built the way H2's own CVE triage and scanner are, with the model kept honest.

3 lessons · ~5 h

Tools The H2 triage service, the H2 scanner, Claude API

1. 01Where AI helps securityTriage, summarization and investigation; where it must not be trusted alone.90 min
2. 02Building a triage serviceAn LLM step that ranks by exploitability; H2's CVE triage as the pattern.120 min
3. 03Keeping it honestGuarding against a crafted input gaming the model; human review where it counts.90 min

Course 05

Adversarial machine learning

Attacks on the models themselves: evasion, poisoning, extraction and inference, and the defenses that raise the cost of each.

3 lessons · ~5 h

Tools ML tooling, adversarial libraries

1. 01Evasion and poisoningFooling a model at inference and corrupting it at training; both demonstrated.120 min
2. 02Extraction and inferenceStealing a model and inferring its training data; the privacy angle.90 min
3. 03DefensesWhat raises the cost of each attack, and what does not work despite the hype.90 min

Course 06

The model supply chain

A model is a dependency you rarely inspect. Provenance, integrity and the poisoned-weights risk, secured the way the DevSecOps path secures code.

3 lessons · ~5 h

Tools Model signing, provenance, the DevSecOps toolchain

1. 01Models as dependenciesWhat you inherit when you pull weights; the trust you extend without noticing.90 min
2. 02Provenance and integritySigning and verifying models; detecting tampering before deploy.120 min
3. 03Poisoned weightsThe backdoored-model risk and the checks that catch it.90 min