Threat Modeling for Agentic AI Systems with Case Study

Show how agentic AI systems act autonomously to fulfill human-provided goals, contrasting with traditional rule-based AI and illustrating with DevOps and loan-processing examples.

Explore key characteristics of agentic AI systems: goal orientation, autonomous decision making, and planning with internal models. Adaptability and self-initiated actions enable them to adjust strategies and trigger subprocesses.

Explain how a single agent architecture enables an autonomous AI system to perceive goals, recall memory, plan, search, evaluate, and act with tools.

Explore the seven layers of agentic AI systems from foundational models to the agent ecosystem, and learn how each layer enables sensing, thinking, and actions with threat modeling and security.

Explore the data layer of agentic ai systems as the foundation for decisions, and identify threats like data exfiltration, data tampering, denial of service, and Rag risks.

Explore the agent framework layer (layer three) of agentic AI systems, including toolkits and SDKs for data integration, and identify threats like backdoors, compromised components, input validation, supply-chain attacks.

Explore deployment layer of agentic ai systems, cloud and on-premise setups, threats like compromised container images, resource hijacking, orchestration attacks, infrastructure as code manipulation, denial of service, and lateral movement.

Examine the evaluation and observability layer of agentic AI systems to monitor performance and detect bias, hallucinations, and compliance risks, plus threats like tampering and data leakage.

Explore the security and compliance layer of agentic ai systems, using ai agents to monitor logs, hunt threats, and enforce policies, while addressing data poisoning, evasion, bias, and transparency threats.

Explore the layer seven agent ecosystem of agentic AI systems, where agents interact with real-world applications, data, and APIs. Identify threats including compromised agents, agent-goal manipulation, malicious discovery, integration risks.

Apply the Maestro threat modeling framework for agentic ai systems to identify adversarial attacks, prompt injection, and goal manipulation, and to prioritize risks with a risk-based approach in multi-agent environments.

Apply a four-step threat modeling process for an agentic AI system by gathering requester information, reviewing architecture, mapping data flows, and reporting threats with mitigations.

Explore threat modeling for an agentic ai system case study by mapping data flow with a dfd and applying manual and automated threat identification across non-genetic and agentic components.

Identify agentic AI threats using the manual approach, examining each layer of agentic AI components to detect prompt injection, adversarial input manipulation, data poisoning, data hallucination, and external service integrations.

Navigate threat modeling for an agentic AI system through a case study. Define project scope, map architecture and data flow, identify threats by severity, and plan remediations and executive feedback.

Identify and mitigate high-severity threats in an agentic AI system using maestro, focusing on prompt injection, adversarial inputs, memory poisoning, and external service risks with layered security.

Explore what a model is in machine learning through a cake-baking analogy, covering data collection, an algorithm, training, refinements, and the final model.

Understand tokens and embeddings as the foundation of ai processing, including tokenization types (word, subword, character), converting tokens to embeddings, and feeding vectors to neural networks to generate outputs.

Explore how ChatGPT relies on transformer-based GPT models trained through unsupervised, supervised, and reinforcement learning from human feedback on a massive Common Crawl dataset.