What question did this study set out to answer?

The aim is to explore and evaluate cross-domain techniques for detecting prompt injection attacks in various datasets.

April 22, 2026Open Access

Beyond Pattern Matching: Seven Cross-Domain Techniques for Prompt Injection Detection

Key Points

The aim is to explore and evaluate cross-domain techniques for detecting prompt injection attacks in various datasets.
Implemented three out of seven techniques across six public datasets for evaluation.
Used local-alignment, stylometric, and fatigue tracking detectors for performance assessment.
Conducted ablation studies with extensive testing to ensure reproducibility and reliability.
Achieved an F1 score increase from 0.033 to 0.378 using the local-alignment detector.
Stylometric detector improved F1 by 11.1 percentage points for indirect injections.
Fatigue tracker effectively blocked probing campaigns with enhanced detection.

Abstract

v2.0: Three of seven cross-domain techniques are implemented and evaluated across six public datasets (deepset/prompt-injections, NotInject, LLMail-Inject, AgentHarm, AgentDojo, and a synthetic indirect-injection benchmark). The local-alignment detector (bioinformatics) lifts F1 on deepset from 0.033 to 0.378 with zero additional false positives. The stylometric detector (forensic linguistics) adds 11.1 percentage points of F1 on indirect injections. The fatigue tracker (materials science) blocks probing campaigns via per-source EWMA hardening. Four-configuration ablation with 868 passing tests and reproducible benchmark harness. All code released under Apache 2.0.

Read Full Paperexternally

AI से पूछें

Bookmark

View Full Paper

Cite This Study

thamilvendhan Munirathinam (Mon,) studied this question.

synapsesocial.com/papers/69e867136e0dea528ddeb631 https://doi.org/https://doi.org/10.5281/zenodo.19664489

AI से पूछें

Bookmark

View Full Paper