All courses

Advanced · module A14

AI Security, Safety and Governance

By the end: Threat-model an AI system, test it adversarially, apply safety techniques, and meet regulatory requirements

Covers
Attacking and defending AI systems, aligning them with human intent, and governing them responsibly
Contents
3 chapters, 13 topics
Time
About 3 hours
Level
Advanced
Your progress
0 / 13

This module builds on Large Language Models: Training, Adaptation and Evaluation. We suggest looking at it first, but the choice is yours.

Chapter 1 Security

  1. 1.1

    Threat Modelling for AI Systems

    Assets, attackers and attack surfaces

    Template: a threat model

    12 minBeing written
  2. 1.2

    Adversarial Examples

    Small changes that fool models, and defences

    Code: an adversarial attack

    15 minBeing written
  3. 1.3

    Data Poisoning and Model Theft

    Attacks on training data and on models themselves

    Case study discussion

    12 minBeing written
  4. 1.4

    Prompt Injection and Jailbreaks

    Attacks on LLM applications and layered defences

    Lab: red-team an assistant

    15 minBeing written
  5. 1.5

    Privacy Attacks and Privacy-Preserving ML

    Membership inference, differential privacy and federated learning

    Code: train with differential privacy

    15 minBeing written

Chapter 2 Safety and Alignment

  1. 2.1

    The Alignment Problem

    Specification gaming, reward hacking and Goodhart's law

    Case studies

    12 minBeing written
  2. 2.2

    Alignment Techniques

    RLHF, constitutional methods and scalable oversight

    Reading and discussion

    12 minBeing written
  3. 2.3

    Mechanistic Interpretability

    Features, circuits and sparse autoencoders

    Code: probe a small model

    15 minBeing written
  4. 2.4

    Red-Teaming and Dangerous-Capability Evaluations

    How frontier models are tested before release

    Lab: design an evaluation

    12 minBeing written

Chapter 3 Fairness and Governance

  1. 3.1

    Fairness Metrics and Mitigation

    Demographic parity, equalised odds and the trade-offs between them

    Code: measure and mitigate unfairness

    15 minBeing written
  2. 3.2

    AI Regulation and Standards

    India's guidelines and DPDP Act, the EU AI Act, NIST AI RMF and ISO/IEC 42001

    Exercise: a compliance checklist

    12 minBeing written
  3. 3.3

    Responsible Scaling and Incident Response

    Safety policies, incident reporting and accountability

    Case study discussion

    10 minBeing written
  4. 3.4

    Module Review and Assessment

    Practical tasks and a scored quiz

    Graded audit

    20 minBeing written