Preprint proposes neurosymbolic defense against prompt injection in AI security operations

An arXiv preprint proposes a neurosymbolic defense-in-depth architecture for security operations centers that use large language models, targeting indirect prompt injection through log poisoning. The authors describe adversaries embedding malicious payloads in system logs to hijack an LLM's operational logic in multistep "promptware" kill chains. The proposed primary layer uses customized SIEM decoders as a deterministic pre-filter to sanitize volumetric padding and signature-based injections at ingestion; a secondary layer uses NeMo Guardrails to enforce semantic boundaries through self-checking validation on structured SIEM alerts before LLM processing. A closed-loop telemetry system provides human-in-the-loop visibility into thwarted attacks in the SOC dashboard. The authors report an experimental evaluation mapped to the MITRE ATLAS taxonomy against diverse prompt injections.