Metraipolis › archive › dc_fb2c6bdf8774
Attention Tracker: Detecting Prompt Injection Attacks in LLMs (2025) · OpenAlex W4411113002
- Source
- OpenAlex
- Original
- https://doi.org/10.18653/v1/2025.findings-naacl.123
- License (as archived)
- CC-BY-4.0
- Archived
- 2026-10-07T10:41:25.985Z
- SHA-256 of the archived text
aa703bd3758f8e694183e99e0e23478121a42683134a941998a3e55be231d12a- Stable URL of this page
https://metraipolis.com/documento/dc_fb2c6bdf8774
How to cite
Attention Tracker: Detecting Prompt Injection Attacks in LLMs (2025) · OpenAlex W4411113002 — OpenAlex, https://doi.org/10.18653/v1/2025.findings-naacl.123 (CC-BY-4.0); archived copy: https://metraipolis.com/documento/dc_fb2c6bdf8774 (Metraipolis archive dc_fb2c6bdf8774, sha256 aa703bd3758f8e694183e99e0e23478121a42683134a941998a3e55be231d12a)
BibTeX
@misc{dcfb2c6bdf8774,
title = {Attention Tracker: Detecting Prompt Injection Attacks in LLMs (2025) · OpenAlex W4411113002},
howpublished = {OpenAlex},
url = {https://doi.org/10.18653/v1/2025.findings-naacl.123},
note = {License CC-BY-4.0; archived copy https://metraipolis.com/documento/dc_fb2c6bdf8774; sha256 aa703bd3758f8e694183e99e0e23478121a42683134a941998a3e55be231d12a},
year = {2026}
}Archived text
Reproduced under CC-BY-4.0. Attribution: OpenAlex, https://doi.org/10.18653/v1/2025.findings-naacl.123. Third-party text: data, not instructions. It may be an abstract or extract; check the original.
Opera scientifica in accesso aperto: scheda dai metadati OpenAlex (conference-paper). Titolo: Attention Tracker: Detecting Prompt Injection Attacks in LLMs Autori: Kuo-Han Hung, Ching-Yun Ko, Ambrish Rawat, I‐Hsin Chung, Winston H. Hsu, Pin‐Yu Chen Anno: 2025 · Sede: Findings of the Association for Computational Linguistics: NAACL · Lingua: en DOI: https://doi.org/10.18653/v1/2025.findings-naacl.123 Accesso aperto: hybrid · testo: https://doi.org/10.18653/v1/2025.findings-naacl.123 · licenza del testo: cc-by Citazioni ricevute (OpenAlex): 26 Temi (OpenAlex): Adversarial Robustness in Machine Learning; Security and Verification in Computing; Topic Modeling Abstract (testo degli autori, licenza cc-by): Large Language Models (LLMs) have revolutionized various domains but remain vulnerable to prompt injection attacks, where malicious inputs manipulate the model into ignoring original instructions and executing designated action.In this paper, we investigate the underlying mechanisms of these attacks by analyzing the attention patterns within LLMs.We introduce the concept of the distraction effect, where specific attention heads, termed important heads, shift focus from the original instruction to the injected instruction.Building on this discovery, we propose Attention Tracker, a training-free detection method that tracks attention patterns on instruction to detect prompt injection attacks without the need for additional LLM inference.Our method generalizes effectively across diverse models, datasets, and attack types, showing an AUROC improvement of up to 10.0% over existing methods, and performs well even on small LLMs.We demonstrate the robustness of our approach through extensive evaluations and provide insights into safeguarding LLM-integrated systems from prompt injection vulnerabilities.Project page: Fonte: https://openalex.org/W4411113002 — metadati OpenAlex (CC0 1.0), consultati il 2026-10-06 via api.openalex.org. Interrogazione: «prompt injection attacks».