Securing LLM systems: AI red teaming and guardrails

How AI systems get attacked and how to defend them, with live red teaming and what the AI Act, Cyber Resilience Act and NIS2 require.

  • Audience:Technical
  • Duration:16 hours · 2 days
  • Format:On-site or online
  • Language:Italian or English

What you'll learn

  • Recognise prompt injection, jailbreaks and data exfiltration
  • Run an AI red teaming session with reproducible results
  • Design layered guardrails on input and output
  • Limit the damage an agent can do with least privilege
  • Turn AI Act, CRA and NIS2 requirements into tests and processes

Course content

4 modules · 16 hours in total

1. Attack surface4 hours
  • Direct and indirect prompt injection
  • Jailbreaks and agent tool abuse
  • Data and model poisoning
2. Live AI red teaming4 hours
  • Method and test plan
  • Attacks on a sample system
  • Report and fix priorities
3. Defence in depth4 hours
  • Input classifiers, output filters
  • Secrets management and personal data redaction
  • Logging and abuse monitoring
4. Regulation and processes4 hours
  • Robustness and cybersecurity in the AI Act
  • Vulnerability and incident reporting under the CRA
  • Playbooks for when something goes wrong

Requirements

  • Software development or IT security experience
  • Basic knowledge of language models

Description

An LLM-based system takes natural language instructions from anyone: that is its strength and also its main weakness. Classic defences are not enough, because the attack can come from a document, an email or a web page the agent reads.

The course comes from our AI red teaming and hardening work, the same we use to protect the assistant on this site. Two hands-on days: attack, measure, fix.

Who this course is for

  • Developers building products on LLMs
  • IT security teams
  • CTOs and technical leads subject to the AI Act, CRA or NIS2

Request information

The course is not sold online: we run it for your team. Write to us and we will send you the programme and a proposal.

Who is asking? *

We reply within 24 business hours with the programme and a proposal. No commitment.

More courses in this category