Integrating Large Language Models (LLMs) into Security Operations Centers (SOCs) automates threat intelligence but opens a critical attack surface: indirect prompt injection via log poisoning. Adversaries embed malicious payloads in system logs, launching multi‑step “promptware” kill chains that hijack the LLM’s reasoning logic.
Securing this pipeline requires balancing deterministic defenses, which are cheap but semantically blind, against pure neural evaluations, which are costly, latency‑heavy, and probabilistic.
We propose a neurosymbolic defense‑in‑depth architecture that guarantees end‑to‑end pipeline integrity. The first layer deploys custom SIEM decoders as a deterministic pre‑filter, performing immediate structural sanitization at the ingestion edge to strip volumetric padding and signature‑based injections. The second layer leverages NeMo Guardrails to enforce strict semantic boundaries through self‑checking validation of the structured SIEM alerts before they reach the LLM.
The framework also incorporates a closed‑loop telemetry subsystem that feeds thwarted attack details back to the SOC dashboard, delivering Human‑in‑the‑Loop (HITL) visibility. Experimental evaluation mapped to the MITRE ATLAS taxonomy across diverse prompt‑injection scenarios shows that the synergistic approach dismantles the promptware kill chain, bounding LLM stochasticity with verifiable constraints and providing a resilient, highly observable defense for next‑generation AI‑SOCs.
Review