llm-prompt-injection

Indirect prompt injection in RAG systems occurs when malicious instructions are embedded within retrieved data sources, manipulating the LLM's behavior. This Skill identifies RAG poisoning and web-browsing injection vectors to guide defensive testing strategies.

To test LLM applications for prompt injection, apply structured payloads targeting instruction overrides and role reassignments. This Skill provides categorized techniques and defense-oriented evaluation criteria to observe model responses against active defenses.

Common LLM tool abuse vectors involve manipulating tool-calling functions and MCP security risks to achieve unauthorized data exfiltration. This Skill maps multi-step escalation patterns to help security teams design robust application defenses.

Yes, this prompt injection playbook includes reference mappings of jailbreak patterns. It provides structured playbooks mapping direct and indirect attack techniques to defense-focused evaluation criteria for AI security testing.

Yes, you can use these techniques for bug bounty testing on AI applications. The Skill applies to security testing across development, QA, and bug-bounty contexts, providing practical reference material to inform risk assessment.