AI Security & Guardrails cover

AI Security & Guardrails

Instructor: CampusX

Language: Hinglish

Validity Period: 1095 days

₹2499 20% OFF

₹1999 including 18% GST

AI Security & Guardrails is a comprehensive course that shows you why LLM applications built without security in mind are open to attack, and how to secure them from the ground up.

You will learn how a partial understanding of RAG, agents and multimodal agents leaves applications insecure, because every retriever, tool and input channel you add without thinking about security becomes a new way in for an attacker. Live demos show how a RAG pipeline can be manipulated, how an attacker can turn an agent's tools and actions against you, and how images, files and other non-text inputs open attack paths that text-only defenses miss.

The course starts with what AI security actually means and why it is harder than traditional application security: outputs are non-deterministic, inputs are natural language, and apps are connected to tools and data. It then moves on to adversarial attacks and the common types used to manipulate an LLM application. You will learn to map any application with Data Flow Diagrams and examine every component through the capability lens: what it can read, write, execute or reach. Finally, you will combine all of this into a step-by-step Threat Modelling Framework you can repeat to secure any LLM-based application.

A major focus of the course is securing a RAG application end to end through the NovaBank case study, which shows why a RAG that works is not yet safe for a bank. Using Data Flow Diagrams, entry and exit points, trust boundaries and the capability table, you will find every weak point in the pipeline and close each one with its own security layer:

  • PII detection and masking, so customers' personal data never reaches the vector database.
  • Authentication and authorization at production level, with pre-retrieval metadata filtering that blocks restricted documents before the search runs.
  • Input guardrails that stop jailbreaks in the user query.
  • Delimiters and a system prompt that treats retrieved documents as data, to defend against prompt injection.
  • Output guardrails built with NVIDIA NeMo Guardrails, which check every answer for private data, wrong facts or a manipulated response before it reaches the user.

The course then takes the same approach to AI agents, using a doctor-booking agent for a website that lists doctors. The agent searches doctors' pages with a tool, checks free slots and books appointments through an MCP server, and learns from past conversations through a shared memory. Anyone on the internet can use it without logging in. Running the same five-step framework, you will see how write and delete capabilities make an agent more dangerous than a RAG: an attack no longer just changes an answer, it makes the agent take real actions, such as leaking patients' details, steering patients to one doctor or cancelling bookings. You will secure each component with its own safety layer:

  • A jailbreak gate, where a separate guard model reads every message, along with recent turns, inside delimiters and blocks authority claims, requests for other people's data and attempts to override the agent's rules.
  • A guard model on every tool and MCP result, which stops indirect prompt injection hidden in doctors' pages and schedule notes.
  • Memory protection, using custom instructions and a judge that checks every write, so personal data and poisoned "lessons" never reach the shared memory.
  • Output guardrails that check every answer against the platform's policies, such as the fee cap, no promised cures, no links to other sites and no other patients' data.

By the end of the course, you will be able to red team both the secured RAG application and the secured agent with Promptfoo, check which security layers hold up at production level, and use the same threat modelling process on any LLM application you build.

Reviews
Other Courses