Analysis
Context Contamination in LLM Analysis of Network Security Logs: Poison with Passive Prompt Injection and Mitigation Evaluation
Publication date unknown · Discovered arxiv.org
Page published
Publication date unknown · First observed: 3 Nov 2026
Coverage timeline
Single-source analysis — one report is available.
Why it matters
AI SOC agents can mistake attacker-controlled evidence for trusted instructions, putting investigation integrity and the data or tools accessible to those agents at risk.
NeuralTrust explains how attacker-controlled logs, alerts, tickets, and emails can inject instructions into AI SOC agents, potentially concealing intrusions, generating false alarms, or triggering unauthorized commands. The cited LogInject study reports up to 88.2% attack success under baseline conditions and evaluates layered defenses that substantially reduce, but do not eliminate, the risk.