GRADUATE COURSE · SAPIENZA UNIVERSITY OF ROME

AI Data-Driven Security

University Lecturer · Master di II Livello in Data Intelligence e Strategie Decisionali

A graduate course on the security of artificial intelligence systems, delivered within the Master di II Livello in Data Intelligence e Strategie Decisionali at Sapienza University of Rome. Across four lectures, the course maps the AI threat landscape, examines prompt injection and LLM manipulation, attacks on models and training data, and the layered defenses - guardrails, AI red teaming and governance - that make AI systems resilient in production.

3 CFU16 Teaching HoursFour Lectures

Back to Home

COURSE CURRICULUM

Four lectures on the security of AI systems

LECTURE 01

Foundations of AI Security

Students learn how AI systems differ from traditional software and why language models introduce entirely new security challenges.

LEARNING OBJECTIVES

Introduce the security lifecycle of modern AI systems and map the new threat landscape that language models introduce.

KEY CONCEPTS

AI threat landscapeAttack surface of LLMsFoundation modelsRisk modelingSecure AI architectureOWASP Top 10 for LLMs

PRACTICAL APPLICATIONS

Threat modeling and secure architecture design for LLM-based platforms.

SKILLS ACQUIRED

AI threat modelingSecure AI architectureLLM risk assessment

LECTURE 02

Prompt Injection & LLM Manipulation

Practical analysis of how attackers manipulate model behaviour and compromise AI assistants.

LEARNING OBJECTIVES

Analyse how crafted inputs manipulate language model behaviour and how attacks escalate through tools, context and external content.

KEY CONCEPTS

Prompt InjectionIndirect Prompt InjectionJailbreakingSystem Prompt leakageContext manipulationAttacks on RAG pipelines

PRACTICAL APPLICATIONS

Security assessment of AI assistants, agents and RAG pipelines.

SKILLS ACQUIRED

Prompt injection testingLLM abuse-case analysisAI assistant hardening

LECTURE 03

Model Attacks & Data Integrity

Students explore how models can be stolen, corrupted or reverse-engineered and how defensive techniques mitigate these risks.

LEARNING OBJECTIVES

Examine attacks against the model itself - its training data, its weights and its outputs.

KEY CONCEPTS

Data PoisoningTraining Data AttacksModel StealingModel ExtractionMembership InferenceAdversarial Examples

PRACTICAL APPLICATIONS

Protecting model supply chains and sensitive training data.

SKILLS ACQUIRED

Model attack analysisData integrity defenceAdversarial robustness

LECTURE 04

Defending AI Systems

Designing resilient AI systems through layered defenses, monitoring and governance.

LEARNING OBJECTIVES

Design layered defenses that keep AI systems resilient and observable in production.

KEY CONCEPTS

GuardrailsInput & Output filteringAI Red TeamingDetection of malicious promptsSecure RAG pipelinesAI Governance & Responsible AI

PRACTICAL APPLICATIONS

AI red teaming, guardrail deployment and AI governance programmes.

SKILLS ACQUIRED

AI red teamingGuardrail designAI governance