Built with and by Teycir Ben Soltane•
How to Use•FAQ•GitHub•arXiv.org•
Share:
ArXivCSExplorer
☆☆Bookmarks🏆RSSHow to UseFAQ
Back to Paper
cs.CRcs.CL

Local ID: 2604.18248v2

AI Summary: gemma4:e4b

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

By Thamilvendhan Munirathinam

Revision History Timeline

v14/20/2026
4/20/2026

“16 pages, 1 table, 25 references. Code: github.com/mthamil107/prompt-shield”

v25/18/2026
5/18/2026

“v3.0 (18 May 2026): Added Sec. 5.6 with independent evaluation on three peer-reviewed benchmarks (Liu, USENIX Sec 2024; Garak, Derczynski 2024; InjecAgent, ACL Findings 2024). 8,276 unseen attacks; cross-benchmark plateau at 35-45% on subtle indirect injection. Abstract, contributions, Sec. 6, and 6 refs updated”

★ Version indexed in Explorer
v37/28/2026
7/28/2026

“v3.0.2 (28 Jul 2026): Abstract scope-hedges specificity to deepset's 56 benign samples and discloses 10 FPs on NotInject (FPR 0.9%->3.8%; Sec. 5.4). Sec. 5.6 reframed as regex-baseline evaluation (d022/d027/d028 disabled in Table 2). Fixes garbled venue for Zhan et al. (NAACL 2025 Findings, arXiv:2503.00061); minor citation cleanup”

v47/30/2026
7/30/2026

“v4 (31 pp, up from 27): adds Sec. 2.4 concurrent-work map (13 papers), Sec. 4.3 marked implemented (d034 ships in v0.7.3), Sec. 5.7 composed-stack adaptive-attack partial run, Sec. 5.8 50-doc held-out benchmark (d027 1.000->0.000 F1 in isolation, composed engine 0.815 F1), Sec. 7 architectural patterns. Repro tag: v0.7.3. Zenodo DOI 10.5281/zenodo.19644135”

Comparing v3 vs v4

Green = Added • Red = Removed

Title Comparison

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

Authors Comparison

No author changes.

v3 Comment

“v3.0.2 (28 Jul 2026): Abstract scope-hedges specificity to deepset's 56 benign samples and discloses 10 FPs on NotInject (FPR 0.9%->3.8%; Sec. 5.4). Sec. 5.6 reframed as regex-baseline evaluation (d022/d027/d028 disabled in Table 2). Fixes garbled venue for Zhan et al. (NAACL 2025 Findings, arXiv:2503.00061); minor citation cleanup”

v4 Comment

“v4 (31 pp, up from 27): adds Sec. 2.4 concurrent-work map (13 papers), Sec. 4.3 marked implemented (d034 ships in v0.7.3), Sec. 5.7 composed-stack adaptive-attack partial run, Sec. 5.8 50-doc held-out benchmark (d027 1.000->0.000 F1 in isolation, composed engine 0.815 F1), Sec. 7 architectural patterns. Repro tag: v0.7.3. Zenodo DOI 10.5281/zenodo.19644135”

Abstract Word Diff

Current open-source prompt-injection detectors converge on two architectural choices: regular-expression pattern matching and fine-tuned transformer classifiers. Both share failure modes that recent work has made concrete. Regular expressions miss paraphrased attacks. Fine-tuned classifiers are vulnerable to adaptive adversaries: a 2025 NAACL Findings study reported that eight published indirect-injection defenses were bypassed with greater than fifty percent attackattack-success successrates rates.under adaptive attacks. This work proposes seven detection techniques that each port a mechanism from a discipline outside LLM security: forensic linguistics, materials-science fatigue analysis, deception technology, local-sequence alignment from bioinformatics, mechanism design, spectral signal analysis, and taint tracking. ThreeEach areproduces implementeda insignal thearchitecturally prompt-shieldindependent v0.4.1of releaseboth (Apacheregex 2.0)matching and evaluatedtransformer acrossclassification, sixso datasetsthe (deepset/prompt-injections,seven NotInject,compose LLMail-Inject,with AgentHarm,existing AgentDojo).defenses Therather local-alignmentthan detectorreplacing liftsthem. F1Four onof deepsetseven fromare 0.033now toimplemented 0.378in withprompt-shield zerov0.7.3 additional(Apache false2.0): positivesd028 onsequence deepset'salignment, fifty-sixd027 benignstylometric samples;discontinuity, thematerials-fatigue sametracking, detectorand addsd034 tenhoneypot falsetool positivesdefinitions on(new thein 339-samplev4.0). NotInjectA benignfour-configuration setablation (FPRacross 0.9%->3.8%),nine abenchmarks specificity(~10,300 regressionsamples) discussedcovers indeepset, SectionNotInject, 5.4.LLMail-Inject, TheAgentHarm, stylometricAgentDojo, detectorand addsan 11.1independent percentageevaluation pointsagainst ofthree F1peer-reviewed onacademic anbenchmarks indirect-injection(Liu benchmark.USENIX The2024, fatigueGarak, trackerInjecAgent). isThis validatedrevision viaadds aSection probing-campaign5.7 integration(composed-stack test.adaptive-attack partial run, evidence for the composability thesis), Section 5.65.8 reports(50-document independentheld-out evaluationbenchmark: ond027 threecollapses peer-reviewed1.000 benchmarksto (Liu,0.000 USENIXF1 Secin 2024;isolation Garak,but 2024;the InjecAgent,composed ACLengine Findingsrecovers 2024)0.815 totallingF1), 8,276and attacks,Section using7 the(three regex-onlyarchitectural baselinepatterns (d022,extracted d027,from d028prompt-shield disabled).in Gang-of-Four format). All code, data, and reproduction scripts are released under Apache 2.0.
View Full Version History on arXiv