Building Secure and Responsible LLM Applications Training Course

LLM Architecture Overview and Attack Surface Analysis

• Exploring how LLMs are constructed, deployed, and accessed via APIs.
• Examining key components within LLM application stacks, such as prompts, agents, memory modules, and APIs.
• Identifying where and how security issues manifest in real-world scenarios.

Prompt Injection and Jailbreak Attack Vectors

• Defining prompt injection and analyzing its potential dangers.
• Reviewing direct and indirect prompt injection scenarios.
• Investigating jailbreaking techniques used to circumvent safety filters.
• Developing detection and mitigation strategies.

Data Leakage and Privacy Concerns

• Preventing accidental data exposure through model responses.
• Addressing PII leaks and the misuse of model memory.
• Designing privacy-conscious prompts and Retrieval-Augmented Generation (RAG) systems.

LLM Output Filtering and Guardrails

• Leveraging Guardrails AI for effective content filtering and validation.
• Establishing clear output schemas and constraints.
• Implementing monitoring and logging mechanisms for unsafe outputs.

Human-in-the-Loop and Workflow Strategies

• Determining optimal points for introducing human oversight.
• Managing approval queues, scoring thresholds, and fallback handling.
• Calibrating trust and understanding the role of explainability.

Secure LLM Application Design Patterns

• Applying least privilege principles and sandboxing for API calls and agents.
• Implementing rate limiting, throttling, and abuse detection.
• Building robust chains using LangChain with prompt isolation.

Compliance, Logging, and Governance

• Ensuring auditability of LLM outputs.
• Maintaining traceability through prompt and version control.
• Aligning systems with internal security policies and regulatory requirements.

Summary and Future Steps