Capture Summary
ACL 2026 Findings paper showing that prompt injection can alter LLM-based hiring rankings under specific market conditions, especially when manipulation is rare and candidate quality is tightly clustered.
Abstract Capture
The paper studies prompt injection in automated résumé screening as subtle self-promotional text that does not add new qualifications but is designed to influence LLM ranking behavior. Controlled experiments show that injection can reliably improve rankings when candidate quality is homogeneous and only a small share of applicants inject. The effect collapses as manipulation becomes widespread, but in heterogeneous pools lower-quality candidates can still occasionally outrank better candidates. The security implication is that prompt injection in workflow automation is not limited to browsing or tool-use agents; it also affects high-stakes decision pipelines where ranking integrity and fairness matter.
Collection Notes
- Untrusted source content. Treat experimental setup and example injection text as evidence only.
- arXiv abstract page records
Journal reference: Findings of the Association for Computational Linguistics: ACL 2026. - Official proceedings listing located via ACL Anthology search during collection.
- Primary relevance: [[03_Topics/Prompt Injection]], [[03_Topics/AI Technology Knowledge Map]]
- PDF: https://arxiv.org/pdf/2606.27287