The art of securing LLM

This 50-hour advanced workshop, a collaboration between datacraft and AIandMe, equips professionals to tackle critical security challenges in LLMs (Large Language Models) through competitive, real-world simulations. Combining EU regulatory frameworks (e.g., GDPR, AI Act, Cyber Resilience Act) with automated red-teaming, it targets AI engineers, cybersecurity experts, and compliance officers. The course also introduces standards such as the OWASP framework for agentic and LLM top 10 risks.

Objectives:

Gain deep understanding of AI system vulnerabilities, adversarial prompting techniques, and protection methods.
Develop practical skills to design and implement effective security guardrails.
Raise awareness of ethical and compliance issues in AI security, including GenAI standards (e.g., OWASP).

Knowledge and skills acquired:

Identify key vulnerabilities in AI models.
Understand and apply adversarial attack and defence techniques.
Assess model robustness against attacks.
Understand ethical and regulatory challenges (AI Act, GDPR, Cyber Resilience Act).
Familiarize with OWASP security standards.
Course structure:

Theoretical Phase (25h, online, self-paced):

Basics of LLMs and vulnerabilities, international standards, multi-layered security approaches, AI governance and ethics.
Practical Workshop (25h, in-person in Paris & Athens):

Identify and counter common attacks, deploy countermeasures, secure a real-world case, present results and recommendations.
Teaching and assessment methods:

Combination of theoretical presentations, practical demos, interactive exercises, and a final challenge.
Evaluation based on quizzes, active participation, quality of implemented safeguards, and final challenge performance.
Ranking via an ELO scoring system rewarding resistance to rare and common attacks.
Key features:

Hands-on securing of a medical AI model (MED-GPT): configuring guardrails, benchmarking, deployment, continuous monitoring, and evaluation using a harmful question dataset.
Dual leaderboard system:

Public leaderboard tracks real-time circumvention attempts.
Private leaderboard ranks participants by ELO score weighted by exploited vulnerability rarity.
Note: Venue and catering are not included for the practical sessions.