Researchers analyzed 196,682 anonymized resumes from the hireEZ platform, as presented at the USENIX Security 2026 conference. The study identified hidden prompt injections in approximately 1% of the documents. These injections consist of text invisible to human readers but accessible to automated parsing systems.
The findings were categorized into two types: explicit instructions designed to influence model behavior and hidden data like additional keywords or skills. Notably, over 90% of the detected injections lacked direct commands, meaning simple filters targeting phrases like 'ignore previous instructions' are insufficient to block them.
The study did not confirm whether these prompts successfully altered hiring outcomes, as the dataset lacked information on recruitment results. Researchers did not test the effectiveness of these attacks on live hiring pipelines, noting that some hidden content might originate from resume templates rather than malicious intent.
A separate report from Fast Company in July highlighted an instance where a hiring manager discovered tiny white text in job applications. This case illustrates the discrepancy between what a human sees and what an AI model processes, which remains a significant security challenge for automated recruitment tools.
