AI-Security for Developers
AI features fail in ways classic AppSec never taught: untrusted text becomes control flow, retrieval pulls in data you didn't vet, and agents act with your credentials. This hands-on course gives developers the mental model and the patterns to build RAG, chat and agent systems that hold up under adversarial use.
What your team can do after
Concrete, transferable skills — not a vendor demo.
Threat-model AI features
Map the attack surface of an LLM / RAG / agent feature before building it — so risk is a design input, not an incident.
Defend against prompt injection
Contain direct and indirect injection — including instructions hidden in retrieved or poisoned content.
Contain blast radius
Design retrieval and tool boundaries — source trust, tenant isolation, least-privilege tools — so one bad input can’t cascade.
Get AuthZ for AI right
Constrain what the model may read and do to the user’s real permissions — mapped onto the same decision model as jSentinel.
Catch regressions in CI
Write adversarial evaluations and red-team prompts that fail the build when a change reopens a hole.
Agenda
Six modules, each ending in a hands-on lab — tailored to your stack and use case.
- The new attack surface
Why text = control flow; the LLM / RAG / agent threat map.
Lab: threat-model a sample assistant. - Prompt injection
Direct & indirect (poisoned retrieval), and how to contain it.
Lab: break and harden a RAG prompt. - Data exfiltration & leakage
System-prompt & secret leakage, cross-tenant bleed.
Lab: plug a leakage path. - Retrieval & tool guardrails
Source trust, tenant isolation, least-privilege tools, confirmations.
Lab: sandbox an over-eager tool. - AuthZ for AI
Constrain what the model may read & do — maps to jSentinel.
Lab: gate an agent action by permission. - Evaluations
Adversarial test suites, red-team prompts, CI wiring.
Lab: build an injection eval suite.
Who it's for
Developers and tech leads adding RAG, chat, copilots or agents to a product — anyone who has to ship AI features and own the consequences.
Prerequisite: general development experience; no ML background required.
Logistics
📋 Course details
- Formats
- On-site · live-online · on-demand (in preparation)
- Duration
- Typically 1 day — tailored to your stack & use case
- Languages
- English or German
- Trainer
- Sven Ruppert