Prompt Injection Attacks Found in 1% of LLM Resume Screening
A systematic study of prompt injection attacks in LLM-based resume screening analyzed approximately 200,000 real-world resumes from hireEZ. Researchers designed tailored detection methods achieving high precision on small-scale validation, outperforming general-purpose detectors. Applying the detector to the full dataset revealed that about 1% of resumes contain prompt injection attempts. This is the first large-scale measurement of such attacks in a real-world application, highlighting a tangible security vulnerability.
Key facts
- First systematic study of prompt injection in LLM-based resume screening
- Analyzed approximately 200,000 real-world resumes from hireEZ
- Tailored detection methods achieved high precision
- Outperformed state-of-the-art general-purpose detectors
- Approximately 1% of resumes contained prompt injection attacks
- Study conducted by researchers using arXiv paper 2605.28999
- Data collected over multiple years
- Highlights real-world prevalence of prompt injection vulnerabilities
Entities
Institutions
- arXiv
- hireEZ