Certified - AI Security Audio Course

This episode examines supply chain and artifact security, focusing on how external dependencies and stored components create systemic risks in AI systems. Artifacts include datasets, model weights, configuration files, and container images, each of which must be treated as high-value assets. For certification purposes, learners must be able to define supply chain risk in the AI context and explain...

This episode introduces MLOps and serving security, focusing on practices that protect the deployment, operation, and continuous delivery of AI models. MLOps extends DevOps principles to AI, requiring controls for model registries, CI/CD pipelines, and serving infrastructure. For certification purposes, learners must know definitions such as model registry, rollback, and shadow deployment, and und...

This episode examines cost and resource abuse, where adversaries or careless users exploit AI systems to drive up compute expenses or deny service to legitimate customers. For exams, learners must understand the concept of denial-of-wallet, token-based denial-of-service, and quota storms that can overwhelm infrastructure. These risks are unique to AI because of resource-intensive inference workloa...

This episode addresses abuse and fraud detection in AI applications, focusing on how adversaries exploit systems for spam, phishing, or marketplace manipulation. For certification purposes, learners must understand definitions of abuse, such as misuse of generative models for disallowed tasks, and fraud, defined as deceptive actions for financial or reputational gain. The exam relevance lies in re...

This episode explores telemetry and observability, emphasizing their importance in detecting anomalies, intrusions, and misuse in AI systems. Telemetry refers to the data collected from inputs, outputs, and system components, while observability describes the ability to interpret that data to understand system health and behavior. For exams, learners must know what to log, how to ensure logs are p...

This episode examines evaluations and test pipelines as essential processes for maintaining AI system security and reliability. Evaluations, or “evals,” are structured tests that measure a model’s behavior against known benchmarks or adversarial scenarios, while pipelines provide the automated flow of regression testing, scorecards, and service-level objectives. For certification purposes, learner...

This episode introduces red teaming as a structured method for probing generative AI systems for vulnerabilities, emphasizing its importance for both exam preparation and real-world resilience. Red teaming involves adopting an adversarial mindset to simulate attacks such as prompt injection, data leakage, or abuse of system integrations. For learners, understanding red team goals, rules of engagem...

This episode examines output validation and policy enforcement as mechanisms for controlling what AI systems produce before results are delivered to users or downstream processes. Output validation ensures that responses conform to expected formats or structures, such as JSON schemas, while policy enforcement applies organizational rules that block disallowed or unsafe outputs. For exam purposes,...

This episode explores authentication (AuthN) and authorization (AuthZ) for large language model (LLM) applications, highlighting their importance in managing identities and permissions. Authentication verifies that a user or system is who they claim to be, while authorization defines what actions or resources they are allowed to access. For certification readiness, learners must understand the dif...

This episode addresses secrets and credential hygiene, emphasizing their critical role in preventing leaks and privilege misuse in AI systems. Secrets include API keys, tokens, passwords, and configuration values embedded in prompts or environments. Learners preparing for exams must understand that secrets frequently appear in AI workflows, often stored insecurely or accidentally revealed in logs...

This episode introduces AI agents as a new and growing attack surface, highlighting how their autonomy and tool integration create unique risks. Agents differ from single-response models by persisting through plan-and-act loops, chaining multiple steps, and invoking external tools or APIs. For certification purposes, learners must understand that these design features expand the system boundary, e...

This episode continues exploration of RAG security by examining context filtering and grounding as defenses for reliable outputs. Learners must understand context filtering as the screening of retrieved documents before they are passed to a model, ensuring that malicious or irrelevant content is excluded. Grounding is defined as aligning model outputs to trusted sources, improving accuracy and red...

This episode explores retrieval-augmented generation (RAG) security, focusing on retrieval and index hardening as foundational defenses. RAG combines language models with external document retrieval, which improves factual grounding but introduces risks. Learners preparing for exams must understand how poisoning of indexes, adversarial queries, and tampered retrieval sources can compromise model o...

This episode introduces adversarial evasion, a class of attacks in which maliciously crafted inputs cause AI systems to misclassify or behave incorrectly. For exam purposes, learners must be able to define adversarial examples, explain why they are often imperceptible to humans, and distinguish them from poisoning attacks, which occur during training. Evasion attacks take place at inference time a...

This episode addresses model theft and extraction, highlighting how adversaries can replicate or steal valuable AI models. Model theft occurs when proprietary weights or architectures are exfiltrated, while model extraction involves querying an exposed API repeatedly to reconstruct decision boundaries or functionality. For exam purposes, learners must be able to distinguish between these two conce...

This episode explores privacy-preserving techniques designed to reduce the risk of sensitive information exposure in AI systems while maintaining utility of the models. Learners must understand concepts such as anonymization, pseudonymization, and data minimization, which limit identifiable information in training sets. Differential privacy is introduced as a mathematical framework that injects st...

This episode introduces privacy attacks in AI systems, focusing on techniques that reveal sensitive or personal information from training data or model behavior. Learners must be able to define key attack types, such as membership inference—determining whether a specific record was included in training—and model inversion, where attackers reconstruct approximate training inputs. The exam relevance...

This episode covers training-time integrity, focusing on the assurance that data, processes, and infrastructure used in model development remain uncompromised. Learners preparing for exams must understand that threats at this stage include data tampering, corrupted labels, or manipulated hyperparameters. Unlike inference-time attacks, which target deployed models, training-time compromises affect...

This episode introduces data poisoning as a high-priority threat in AI security, where adversaries deliberately insert malicious samples into training or fine-tuning datasets. For exam readiness, learners must understand how poisoning undermines model accuracy, introduces backdoors, or biases outputs toward attacker goals. The relevance of poisoning lies in its persistence, as compromised models m...

This episode explains the distinction and overlap between content safety and security in AI systems, a concept often emphasized in both professional practice and certification exams. Content safety refers to filtering or moderating outputs to prevent harmful or offensive material, while security focuses on protecting systems and assets from adversarial manipulation or data loss. Although they are...

This episode examines indirect and cross-domain prompt injections, which expand the attack surface by embedding malicious instructions in external sources such as documents, websites, or email content. Unlike direct injection, where the attacker provides inputs to the model directly, these threats exploit retrieval or integration features that feed information into the AI system automatically. Lea...

This episode introduces prompt injection and jailbreaks as fundamental AI-specific security risks. It defines prompt injection as malicious manipulation of model inputs to alter behavior and describes jailbreaks as methods for bypassing built-in safeguards. For certification purposes, learners must understand these concepts as new categories of vulnerabilities unique to AI, distinct from but conce...

This episode examines data lifecycle security, covering the journey of data from collection and labeling through storage, retention, deletion, and provenance management. It explains why data is the foundation of AI system reliability and how its misuse or compromise undermines security objectives. For certification preparation, learners are introduced to key definitions of provenance, integrity, a...

This episode explains the architecture of AI systems, breaking down their stages and components to show how trust boundaries shift across the lifecycle. Training, inference, retrieval-augmented generation (RAG), and agent frameworks are introduced as discrete but interconnected environments, each with distinct risks. For exam relevance, learners are expected to identify these architectural element...

This episode defines the AI security landscape by mapping the assets, attack surfaces, and emerging threats that distinguish AI from classical application security. It introduces critical components such as training data, model weights, prompts, and external tools, explaining why each must be protected as an asset. The relevance for certification exams lies in understanding how these components sh...