Agentic AI Security & LLM Governance Career Bootcamp

AI security builds on traditional security by adding AI-specific controls, since prompts, documents, and chat history shape model behavior.

Map the AI threat landscape across prompts, documents, and tool calls. Apply layered defenses: prompt validation, source trust in RAG, and restricted tool use plus output checks.

Implement layered ai security by validating prompts, isolating content, securing retrieval for RAG systems, and limiting tools, creating an observable pipeline with checks, audits, and monitoring.

Explore how prompt injection and roleplay attacks reveal weaknesses in chat assistants, and learn to implement app-level prompt validation, risk scoring, and audit logs for local and OpenAI models.

Explore direct prompt injection and its security risks in AI apps, and learn input validation, guardrails, and prompt hierarchy to protect model behavior.

Explore prompt hierarchy across system, developer, and user messages to secure ai applications, prevent prompt injection, and implement validation, guardrails, and monitoring for safe llm governance.

Implement a prompt validation layer to prevent injection attempts by normalizing messages, checking blocked patterns, and returning a safe block message before the model call in a streamlit chat app.

Add a basic guardrail layer that blocks prompts and returns a consistent safe refusal before the model, with a dedicated guardrails file and a consistent refusal message.

Test and strengthen prompt injection defenses by running a local app in a virtual environment, validating prompts with guardrails that block risky inputs before reaching the model.

Develop and integrate a prompt risk analyzer that scores jailbreak signals—such as roleplay, encoding, hidden instructions, and multi-turn manipulation—before generating a response.

Discover how retrieval augmented generation (RAG) grounds model answers by retrieving relevant documents and embedding chunks for vector search. Explore the five-step rx pipeline and essential security considerations.

Design and implement a document-based retrieval-augmented generation assistant from scratch, within the agentic ai security and llm governance bootcamp framework.

Explore how poisoned retrieved context can corrupt the source document, and learn defenses like trusted source validation, metadata checks, context isolation, and sanitization.

Implement trusted source validation to gate document ingestion. Enforce approved routes, normalized paths, and file type and size limits before text extraction, chunking, or embedding.

Add context isolation to the RAG assistant by wrapping retrieved content with start and end markers, treating untrusted text as reference data and separating it from trusted instructions.

Apply content sanitization to retrieved text by adding a sanitizer module that neutralizes risky prompt injection patterns before they enter the model context. Support trusted source validation and context isolation.

Add an enforcement layer to the rag pipeline by building a reusable injection detector and a policy that blocks unsafe chunks before they reach the model, with a safe fallback.

Test and demonstrate a tool calling chat assistant, validating tool usage and execution with a calculator, weather, and email draft tools. Examine vulnerabilities and defenses to strengthen tool security.

Examine the weaknesses of a tool calling chat assistant and how unauthorized tool calls can bypass authorization, with emphasis on missing parameter validation and the need for human-in-the-loop approvals.

Add a human approval gate for sensitive tool actions, pausing proposals until explicit sign-off. Build a reusable approval policy integrated with registry, service layers, and a user interface.

Verify tool security by running the app and testing that unsafe inputs are blocked, sensitive tools require approval, and the system validates tool parameters before execution.

Enable persistent memory with a local JSON memory store, formatted for prompts and injected into tool and chat flows, underpinning security-focused memory validation and approval.

Demonstrates memory retrieval by storing an approved user preference and showing its visibility and control, then retrieves it to personalize responses, while noting stored context can be dangerous if unvalidated.

This lecture demonstrates memory poisoning, showing how malicious preferences stored in memory can influence future answers and security recommendations, highlighting memory as untrusted input and need for validation and approval.

Add a dedicated memory validator to safely gate memory writes, using regex patterns and length checks to prevent memory poisoning, and surface policy rejections through the service layer.

Introduce a human approval gate before persisting memory, coupling validation with an approval policy to pause, review, and approve proposed memories in a two-step storage flow.