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
Artificial Intelligence Risk Environment
- Distinguishing characteristics of AI security: inherent non-determinism, opaque decision-making processes, and the expansion of the attack surface through prompt interfaces
- Classification of threats: categorizing risks across training, inference, and supply chain phases
- Adversary profiling: identifying potential threat actors and their motivations targeting AI infrastructure
NIST and OWASP Frameworks for Large Language Models
- Prompt injection methodologies: analyzing direct and indirect exploitation vectors
- Vulnerabilities in output processing: mitigating insecure data handling and cross-plugin request forgery
- Data integrity risks: addressing training data contamination and supply chain compromises
- Service availability and confidentiality: preventing model denial of service, protecting sensitive data, and limiting excessive system agency
- Practical assessment: conducting controlled tests against designated applications to validate each vulnerability category
Adversarial Testing and Red Team Operations
- Injection technique taxonomy: evaluating direct, indirect, multi-turn, and multi-modal attack strategies
- Automated threat discovery: leveraging tools such as Giskard, Garak, and custom fuzzing utilities for systematic testing
- Jailbreak assessment: classifying evasion attempts and evaluating defensive postures
- Operationalizing red teams: developing harnesses for continuous security validation of LLM deployments
Threat Vectors Against Model Architecture
- Model extraction attacks: reconstructing model weights and functionality through systematic API queries
- Membership inference risks: determining whether specific records were included in the training dataset
- Adversarial perturbation techniques: manipulating inputs to deceive classifiers and embedding systems
- Data contamination: introducing corrupt data to establish backdoors or degrade system performance
Input Validation and Output Management
- Advanced input sanitization: extending beyond conventional web application security controls
- Output mitigation strategies: filtering for toxicity, personally identifiable information (PII) exposure, and unauthorized code execution via hallucinations
- Implementation of guardrail infrastructure: utilizing platforms such as NeMo and Guardrails AI alongside custom policy enforcement
- Enforcement of structured outputs: establishing strict data schemas as a critical security boundary
AI Supply Chain Integrity
- Model provenance verification: ensuring the authenticity and integrity of foundational models
- Dependency management: scanning ML frameworks and model artifacts for known vulnerabilities
- Secure deployment practices: implementing sandboxing, network isolation, and least-privilege access protocols
- Third-party model vetting: screening fine-tuned and community-sourced models for embedded malicious code
AI System Operational Security
- Access management: securing model endpoints, vector databases, and agent tooling
- Audit and monitoring: maintaining comprehensive logs for all model interactions and automated decisions
- Incident response protocols: addressing AI-specific breaches where the model logic is compromised
- Continuous integration security: embedding AI-focused testing within ML development pipelines
Establishing an AI Security Program
- Maturity assessment: defining a roadmap for AI security capabilities aligned with organizational needs
- Program integration: incorporating AI controls into existing application security and cloud infrastructure frameworks
- Regulatory compliance: navigating governance structures and emerging standards for AI systems
- Policy development: creating and maintaining an organizational playbook for AI risk management, for government and private sector stakeholders alike
Requirements
* Demonstrated capability in the operational deployment of machine learning models and large language model applications.
* Competence in core security principles, including identity management, access control mechanisms, and risk assessment methodologies.
* Proficiency in Python programming to facilitate adversarial validation and testing procedures.
**Target Participants**
* Security professionals transitioning to address emerging threats within artificial intelligence and machine learning ecosystems.
* Data science personnel tasked with ensuring the integrity and resilience of predictive systems.
* Ethical hacking specialists integrating AI-driven technologies into their evaluation frameworks for government operations.
14 Hours
