Catching prompt injections hidden in PDFs
Summary
Catching prompt injections hidden in PDFs examines how AI QA workflows can be manipulated via prompt injections embedded in PDF content. The post details a methodology for testing and defending against such attacks, including a multi-layered “judge” system to detect hidden payloads, techniques for obfuscating text, and practical examples of how PDFs could subvert AI prompts. It also reflects on the challenges of building robust, secure AI QA processes in the face of evolving threat models.