AI Security

Organisations must understand how to secure their AI systems. This in-depth course delves into the AI security landscape, addressing vulnerabilities like prompt injection, denial of service attacks, model theft, and more. Learn how attackers exploit these weaknesses and gain hands-on experience with proven defense strategies and security APIs.

Discover how to securely integrate LLMs into your applications, safeguard training data, build robust AI infrastructure, and ensure effective human-AI interaction. By the end of this course, you'll be equipped to protect your organization's AI assets and maintain the integrity of your systems.

This course will cover the following topics:

• Introduction to AI Security
• Types of AI Systems and Their Vulnerabilities
• Understanding and Countering AI-specific Attacks
• Ethical and Reliable AI
• Prompt Injection
• Model Jailbreaks and Extraction Techniques
• Visual Prompt Injection
• Denial of Service Attacks
• Secure LLM Integration
• Training Data Manipulation
• Human-AI Interaction
• Secure AI Infrastructure

Learning Outcomes

• Gain a comprehensive understanding of AI technologies and the unique security risks they pose
• Learn to identify and mitigate common AI vulnerabilities
• Gain practical skills in securely integrating LLMs into applications
• Understand the principles of responsible, reliable, and explainable AI
• Familiarize themselves with security best practices for AI systems
• Stay updated with the evolving threat landscape in AI security
• Engage in hands-on exercises that simulate real-world scenarios

No prerequisites, aside general understanding of AI principles.

Day 1

Introduction to AI security

• What is AI Security?
  • Defining AI
  • Defining Security
  • AI Security scope
  • Beyond this course
• Different types of AI systems
  • Neural networks
  • Models
  • Integrated AI systems
• From Prompts to Hacks
  • Use-cases of AI systems
  • Attacking Predictive AI systems
  • Attacking Generative AI systems
  • Interacting with AI systems
• What does 'Secure AI' mean?
  • Responsible AI
  • Reliable, trustworthy AI
  • Explainable AI
  • A word on alignment
  • To censor or not to censor
• Exercise: Using an uncensored model
  • Using an uncensored model

Using AI for malicious intents

• Deepfake scam earns $25M
  • You would never believe, until you do
  • Behind deep fake technology
• Voice cloning for the masses
  • Imagine yourself in their shoes
  • Technological dissipation
• Social engineering on steroids
• Levelling the playing field
• Profitability from the masses
• Shaking the fundamentals of reality
• Donald Trump arrested
• Pentagon explosion shakes the US stock market
• How humans amplify a burning Eiffel tower
• Image watermarking by OpenAI
• Exercise: Image watermarking
  • Real or fake?

The AI Security landscape

• Attack surface of an AI system
  • Components of an AI system
  • AI systems and model lifecycle
  • Supply-chain is more important than ever
  • Models accessed via APIs
  • APIs access by models
  • Non-AI attacks are here to stay
• OWASP Top 10 and AI
  • About OWASP and it's Top 10 lists
  • OWASP ML Top 10
  • OWASP LLM Top 10
  • Beyond OWASP Top 10
• Threat modeling an LLM integrated application
  • A quick recap on threat modeling
  • A sample AI-integrated application
  • Sample findings
  • Mitigations
• Exercise: Threat modeling an LLM integrated application
  • Meet TicketAI, a ticketing system
  • TicketAI's data flow diagram
  • Find potential threats

Prompt Injection

• Attacks on AI systems - Prompt injection
  • Prompt injection
  • Impact
  • Examples
  • Indirect prompt injection
  • From prompt injection to phishing
• Advanced techniques - SudoLang: pseudocode for LLMs
  • Introducing SudoLang
  • SudoLang examples
  • Behind the tech
  • A SudoLang program
  • Integrating an LLM
  • Integrating an LLM with SudoLang
• Exercise: Translate a prompt to SudoLang
  • A long prompt
  • A different solution
• Exercise: Prompt injection - Get the password for levels 1 and 2
  • Get the password!
  • Classic injection defense
  • Levels 1-2
  • Solutions for levels 1-2

Day 2

Prompt Injection

• Attacks on AI systems - Model jailbreaks
  • What's a model jailbreak?
  • How jailbreaks work?
• Jailbreaking ChatGPT
  • The most famous ChatGPT jailbreak
  • The 6.0 DAN prompt
  • AutoDAN
• Exercise: Jailbreaking - Get the password for levels 3, 4, and 5
  • Get the password!
  • Levels 3-5
  • Use DAN against levels 3-5
• Tree of Attacks with Pruning (TAP)
  • Tree of Attacks explained
• Attacks on AI systems - Prompt extraction
  • Prompt extraction
• Exercise: Prompt Extraction - Get the password for levels 6 and 7
  • Get the password!
  • Level 6
  • Level 7
  • Extract the boundaries of levels 6 and 7
• Defending AI systems - Prompt injection defenses
  • Intermediate techniques
  • Advanced techniques
  • More Security APIs
  • ReBuff example
  • Llama Guard
  • Lakera
• Attempts against a similar exercise
  • Gandalf from Lakera
  • Types of Gandalf exploits
• Exercise: The Real Challenge - Get the password for levels 8 and 9
  • Get the password!
  • Level 8
  • Level 9
• Other injection methods
  • Attack categories
  • Reverse Psychology
• Exercise: Reverse Psychology
  • Write an exploit with the ChatbotUI
• Other protection methods
  • Protection categories
  • A different categorization
  • Bergeron method
• Sensitive Information Disclosure
  • Relevance
  • Best practices

Visual Prompt Injection

• Attack types
  • New Tech, New Threats
  • Trivial examples
  • Adversarial attacks
• Tricking self-driving cars
  • How to fool a Tesla
  • This is just the beginning
• Exercise: Image recognition with OpenAI
  • Invisible message
  • Instruction on image
• Exercise: Adversarial attack
  • Untargeted attack with Fast Gradient Signed Method (FGSM)
  • Targeted attack
• Protection methods
  • Protection methods

Denial of Service

• Chatbot examples
  • Attack scenarios
  • Denial of Service
  • DoS attacks on LLMs
  • Risks and Consequences of DoS Attacks on LLMs
• Prompt routing challenges
  • Attacks
  • Protections
• Exercise: Denial of Service
  • Halting Model Responses

Model theft

• Know your enemy
  • Risks
• Attack types
  • Training or fine-tuning a new model
  • Dataset exploration
• Exercise: Query-based model stealing
  • OpenAI API parameters
  • How to steal a model
• Protection against model theft
  • Simple protections
  • Advanced protections

Day 3

LLM integration

• The LLM trust boundary
  • An LLM is a system just like any other
  • It's not like any other system
  • Classical problems in novel integrations
  • Treating LLM output as user input
  • Typical exchange formats
  • Applying common best practices
• Exercise: SQL Injection via an LLM
• Exercise: Generating XSS payloads
• LLMs interaction with other systems
  • Typical integration patterns
  • Function calling dangers
  • The rise of custom GPTs
  • Identity and authorization across applications
• Exercise: Making a call with invalid parameters
• Exercise: Privilege escalation via prompt injection
• Principles of security and secure coding
• Racking up privileges
  • The case for a very capable model
  • Exploiting excessive privileges
  • Separation of privileges
  • A model can't be cut in half
  • Designing your model privileges
• A customer support bot going wild
• Exercise: Breaking out of a sandbox
• Best practices in practice
  • Input validation
  • Output encoding
  • Use frameworks

Training data manipulation

• What you train on matters
  • What data are models trained on?
  • Model assurances
  • Model and dataset cards
• Exercise: Verifying model cards
• A malicious model
• A malicious dataset
  • Datasets and their reliability
  • Attacker goals and intents
  • Effort versus payoff
  • Techniques to poison datasets
• Exercise: Let's construct a malicious dataset
• Verifying datasets
  • Getting clear on objectives
  • A glance at the dataset card
  • Analysing a dataset
• Exercise: Analysing a dataset
• A secure supply chain
  • Proving model integrity is hard
  • Cryptographic solutions are emerging
  • Hardware-assisted attestation

Human-AI interaction

• Relying too much on LLM output
  • What could go wrong?
  • Countering hallucinations
  • Verifying the verifiable
  • Referencing what's possible
  • The use of sandboxes
  • Building safe APIs
  • Clear communication is key
• Exercise: Verifying model output

Secure AI infrastructure

• Requirements of a secure AI infrastructure
  • Monitoring and observability
  • Traceability
  • Confidentiality
  • Integrity
  • Availability
  • Privacy
• Privacy and the Samsung data leak
• LangSmith
• Exercise: Experimenting with LangSmith
• BlindLlama