Home / Cybersecurity / AI/LLM Red Teaming
Next batch January 2027. 15 seats.
AI/LLM Red Teaming
Test AI features the way an attacker would. You probe chatbots, retrieval-augmented (RAG) apps and AI agents for prompt injection, data leaks and unsafe actions, then add guardrails and prove they hold. Every exercise runs against practice apps in an isolated environment.
- Level
- 2, after the Foundation or equivalent
- Hours
- 20 hours, 7 modules
- Format
- Live online
- Schedule
- 5-day intensive or 5 weekends
- Fee
- On enquiry, team rates available
What you’ll learn
- Scope an AI red-team exercise, agree the rules of engagement and set up an isolated practice environment.
- Threat-model an LLM application using the OWASP Top 10 for LLM Applications and MITRE ATLAS.
- Test for direct and indirect prompt injection, including instructions hidden in documents and web pages an app reads.
- Find sensitive data and system-prompt leaks in chatbots and RAG apps.
- Test AI agents for excessive permissions and unsafe tool use, and limit what they can do.
- Automate LLM testing with open-source tools and read the results critically.
- Add guardrails, retest, and write findings a product team can act on.
Seven modules, from threat model to retest.
Each module ends in a hands-on project. In the capstone, you red-team a practice AI assistant and deliver a full report.
| Module | Focus | Hands-on project | Hours |
|---|---|---|---|
| 0. Rules and setup | Scope, authorisation, isolated lab | Set up a local model and a practice chatbot | 1 |
| 1. How LLM apps fail | OWASP LLM Top 10, MITRE ATLAS | Threat-model a RAG assistant | 3 |
| 2. Prompt injection | Direct and indirect injection | Hijack a practice assistant through a document it reads | 4 |
| 3. Data leakage | System prompts, retrieval, sensitive data | Find what a RAG app reveals that it shouldn’t | 3 |
| 4. Agents and tools | Excessive agency, unsafe output handling | Make an agent take an action it shouldn’t, then restrict it | 3 |
| 5. Automated testing | Open-source LLM scanners | Run Garak and promptfoo and triage the results | 3 |
| 6. Guardrails and reporting | Defences, retesting, writing findings | Add guardrails, retest and write the report | 3 |
All testing is done on practice apps we provide, never on live services you don’t own. For web and API testing beyond AI features, see Track 1B: Application Security.
Private batches for teams shipping AI
We run this course privately for product, engineering and security teams, built around the AI features you are building. Online or at your office.
Do I need a machine-learning background?
No. You test AI features from the outside, the way a user or attacker meets them. You need to be comfortable with web apps, APIs and short Python scripts.
Do I need paid AI accounts?
No. The labs use open-weight models that run in the practice environment we provide.
Will I get a certificate?
Yes, a Sophorion certificate of completion with an ID your employer can check on our verification page, plus your capstone report.
Reserve a seat in the January batch.
Tell us your background and we’ll confirm the schedule that suits you.
