AI Security, Red Teaming & LLM Defense

• OWASP Top 10 for Large Language Model Applications (LLM01 Prompt Injection to LLM10 Model Theft)
• Direct jailbreak taxonomies: Roleplay persona switches, base64 obfuscation, multi-turn crescendo attacks
• Indirect prompt injection in RAG pipelines: Malicious payloads hidden in PDFs, emails, and web search results

• Deploying NVIDIA NeMo Guardrails (Colang scripts for topical rails, moderation, and fact-checking)
• Llama Guard 3 and Presidio for real-time PII anonymization and toxic response blocking
• Defense-in-depth: Dual-LLM validation architectures (Untrusted Content vs Decision-Maker Model)

• Training data extraction attacks and membership inference defense
• RAG data poisoning: Protecting vector databases from adversarial document injection
• AI governance risk cards: Threat modeling enterprise AI agents and red team reporting

Q1: What is an Indirect Prompt Injection attack and how do you mitigate it in a RAG system?

Indirect Prompt Injection occurs when an attacker places adversarial instructions into an external data source (a website, email, or resume) that the LLM later retrieves via RAG or web search. When the LLM ingests this untrusted content, the hidden prompt overrides the system instructions (e.g. instructing the agent to exfiltrate user data). Mitigation includes: isolating untrusted data in separate context blocks, using dual-model architectures, and deploying input guardrails to scan retrieved chunks before context assembly.