Thank you for sending your enquiry! One of our team members will contact you shortly.
Thank you for sending your booking! One of our team members will contact you shortly.
Duration 14 hours
Course Outline
The AI Threat Environment
- The uniqueness of AI security: dealing with non-determinism, opaque reasoning, and prompts as attack vectors.
- Categorizing attacks: training-time, inference-time, and supply chain threats.
- Understanding the ML adversary model: identifying who attacks AI systems and their motivations.
OWASP Top 10 for LLM Applications
- Prompt injection: analyzing both direct and indirect attack vectors.
- Risks in insecure output handling and cross-plugin request forgery.
- Vulnerabilities stemming from training data poisoning and the supply chain.
- Threats including model denial of service, sensitive data exposure, and excessive agency.
- Practical lab session: exploiting each OWASP category within a test application.
Red Teaming for Prompt Injection and Jailbreaks
- Classifying injection techniques: direct, indirect, multi-turn, and multi-modal approaches.
- Utilizing automated red-teaming tools such as Giskard, Garak, and custom fuzzing solutions.
- Assessing jailbreak classifications and evaluating defense effectiveness.
- Constructing a red-team framework for ongoing LLM security testing.
Attack Vectors and Defenses at the Model Level
- Model extraction: preventing the theft of model weights and functionality via API interactions.
- Membership inference: detecting whether specific data was part of the training set.
- Adversarial examples: understanding perturbations that mislead classifiers and embeddings.
- Data poisoning: guarding against corrupted training data that may introduce backdoors or reduce performance.
Security Controls for Inputs and Outputs
- Advanced input sanitization techniques that go beyond standard web defenses.
- Output filtering to prevent toxicity, PII leakage, and hallucinated code execution.
- Guardrails as core security infrastructure: leveraging NeMo, Guardrails AI, and custom policies.
- Enforcing structured outputs to establish a clear security boundary.
Securing the AI Supply Chain
- Model provenance: ensuring the authenticity and integrity of models.
- Scanning dependencies across ML frameworks and model formats.
- Secure model serving practices, including sandboxing, network isolation, and least-privilege access.
- Auditing fine-tuned and community-sourced models for potential embedded malware.
Operational Security for AI Systems
- Managing access controls for model endpoints, vector stores, and agent tools.
- Implementing comprehensive audit logs for all model interactions and decisions.
- Incident response strategies for AI-specific breaches, including cases where the model itself is compromised.
- Integrating continuous security testing into CI/CD pipelines for ML workflows.
Establishing an AI Security Program
- Defining an AI security maturity model and strategic roadmap.
- Weaving AI security into existing AppSec and cloud security frameworks.
- Navigating governance frameworks and emerging regulations for AI systems.
- Developing and sustaining an organizational AI security playbook.
Requirements
- Practical experience in deploying ML models or LLM applications in production environments.
- A solid understanding of core security principles, including authentication, authorization, and threat modeling.
- Proficiency in Python to support adversarial testing activities.
Target Audience
- Security professionals looking to expand their expertise into AI/ML threat domains.
- ML engineers tasked with ensuring model safety and robustness.
- Red team specialists incorporating AI systems into their testing scope.
Testimonials (1)
inventory and identifying the different risk exposures within AI
