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.
Course Outline
The AI Threat Landscape
- Why AI security differs: non-determinism, opaque reasoning, and prompts as an attack surface.
- Attack taxonomy: distinguishing between training-time, inference-time, and supply chain attacks.
- The ML adversary model: understanding who attacks AI systems and their motivations.
OWASP Top 10 for LLM Applications
- Prompt injection: exploring direct and indirect attack vectors.
- Insecure output handling and cross-plugin request forgery.
- Training data poisoning and supply chain vulnerabilities.
- Model denial of service, sensitive information disclosure, and excessive agency.
- Hands-on lab: exploiting each OWASP category against a test application.
Prompt Injection and Jailbreak Red Teaming
- Taxonomy of injection techniques: direct, indirect, multi-turn, and multi-modal.
- Automated red-teaming using Giskard, Garak, and custom fuzzing tools.
- Jailbreak classification and evaluation of defense mechanisms.
- Constructing a red-team harness for continuous LLM security testing.
Model-Level Attacks and Defenses
- Model extraction: stealing model weights and functionality via API queries.
- Membership inference: determining whether specific data was part of the training set.
- Adversarial examples: perturbations designed to fool classifiers and embeddings.
- Data poisoning: corrupting training data to induce backdoors or degrade performance.
Input and Output Security Controls
- Advanced input sanitization beyond traditional web defenses.
- Output filtering for toxicity, PII leakage, and hallucinated code execution.
- Guardrails as security infrastructure: utilizing NeMo, Guardrails AI, and custom policies.
- Using structured output enforcement as a security boundary.
AI Supply Chain Security
- Model provenance: verifying model authenticity and integrity.
- Dependency scanning for ML frameworks and model formats.
- Secure model serving practices: sandboxing, network isolation, and least-privilege access.
- Vetting fine-tuned and community models for embedded malware.
Operational Security for AI Systems
- Access control for model endpoints, vector stores, and agent tools.
- Audit logging for every model interaction and decision made.
- Incident response protocols for AI-specific breaches, including when the model itself is compromised.
- Implementing continuous security testing in CI/CD for ML pipelines.
Building an AI Security Program
- Establishing an AI security maturity model and roadmap.
- Integrating AI security into existing AppSec and cloud security programs.
- Navigating governance frameworks and emerging regulations for AI systems.
- Creating and maintaining an organizational AI security playbook.
Requirements
- Experience deploying ML models or LLM applications in production environments.
- Familiarity with core security concepts, including authentication, authorization, and threat modeling.
- Proficiency in Python for conducting adversarial testing exercises.
Target Audience
- Security engineers expanding their focus into AI/ML threat surfaces.
- ML engineers responsible for ensuring model safety and robustness.
- Red team members incorporating AI systems into their testing scope.
14 Hours
