indirect-injection-detection

Indirect prompt injection occurs when malicious commands are embedded in external content processed by AI agents, causing goal hijacking. This Skill detects such hidden directives within RAG results, documents, or web pages to prevent unauthorized data leaks.

You can test for hidden instructions by simulating external content with embedded malicious directives. This Skill uses dialogue to simulate fake documents or retrieved chunks, validating whether your AI agent follows user prompts or falls for the injected traps.

Yes, you can detect goal hijacking in document analysis workflows by validating AI adherence to user-defined tasks versus hidden directives. This ensures the agent processes documents securely without executing embedded unauthorized commands.

Yes, this works for web content processing red teaming by testing if the AI follows hidden commands embedded in processed web pages rather than the user's explicit prompt. It validates agent adherence to user-defined tasks against hidden directives.

The primary constraint is that detection relies on simulating external content with fake documents via dialogue to test agent responses. It validates adherence to user-defined tasks but may not anticipate novel or highly complex hidden directive patterns.