Hidden Text in PDFs Can Hijack AI Assistants, Researchers Warn of New Attack Vector

Security researchers have discovered that hidden text within PDF files can be leveraged to hijack AI assistants, coercing them into performing unintended operations. The attack exploits the parsing mechanisms of AI models by embedding malicious instructions in invisible text layers, which are then triggered without the knowledge of users or systems. As reported by Decrypt, this technique has been successfully demonstrated against several mainstream AI assistants, posing risks to automated processes such as document handling and data extraction.

Attackers achieve this by setting text to match the background color or placing it within clipped regions, ensuring the PDF appears visually normal while AI systems still read the hidden content during text extraction. Researchers showcased scenarios where assistants were manipulated into leaking conversation histories or executing specific commands, with some models failing to differentiate between visible and concealed content. While no large-scale exploits have been publicly documented, experts caution that as AI agents gain broader permissions, the threat of such injection attacks is poised to escalate significantly.

This issue underscores the challenge of "input trustworthiness" in AI security. Organizations deploying AI to process external documents must consider validating text sources or restricting agents from autonomously executing sensitive actions. The vulnerabilities have been reported to select model vendors, though a unified fix has yet to emerge.

Source: https://decrypt.co/375269/hidden-text-pdfs-hijacking-ai-assistant-prompt-injection