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

v3.0.0 (30 July 2026) — substantive revision. Consolidates the arXiv v3.0, v3.0.1, v3.0.2, and v4.0 revision cycles. New substantive content: Section 2.4: concurrent-work map covering 13 concurrent 2026 preprints and publications on overlapping problems (WebSentinel, SnapGuard, DataSentinel, ClawGuard, AgentSpec, Prismata, DualView, Token-Flow Firewall, PromptSentinel-X, AgentWatcher, NetInjectBench, AgentSentry, WARD). Section 4.3: honeypot tool definitions marked IMPLEMENTED (d034 ships in prompt-shield v0.7.3, released 30 July 2026). Fourth of seven cross-domain techniques now implemented. Section 5.6: independent evaluation against three peer-reviewed benchmarks (Liu USENIX Security 2024, Garak, InjecAgent), regex-only baseline configuration, ~8,300 unseen attacks. Section 5.7: adaptive-attack evaluation of d028 (Smith-Waterman) and a two-layer composed-stack partial run supporting the composability thesis. Section 5.8: 50-document held-out benchmark of LLM-template-generated business documents with paraphrased indirect-injection payloads. Reports d027 stylometric detector collapsing from 1.000 F1 (paper Section 5.2 synthetic corpus) to 0.000 F1 (this held-out corpus) while the composed engine recovers 0.815 F1. Published against self-interest for scientific integrity. Section 7: three architectural patterns extracted from prompt-shield in Gang-of-Four format — Tool-Result Boundary Gate (independently converged on by four concurrent 2026 systems), Federated Signed Threat Intelligence, Attack-Family Taxonomic Projection. Corrigenda vs v2.0.0: NotInject false-positive rate corrected 0.9% → 3.8% (paper Section 5.4 disclosure); scope-hedged the “zero additional false positives” claim to deepset’s 56 benign samples. Test count 1,155 → 1,232 passing on prompt-shield main. Reproducibility bundle pinned to prompt-shield tag v0.7.3 (github.com/mthamil107/prompt-shield/releases/tag/v0.7.3). Companions: arXiv:2604.18248 v4 (in moderation as of 30 July 2026). Code + benchmarks: pip install prompt-shield-ai==0.7.3 or clone github.com/mthamil107/prompt-shield at v0.7.3. License: paper CC BY 4.0, code Apache 2.0, benchmark data CC0.