Work

AgentShield Hybrid Runtime Firewall Architecture for Secure Tool Governance in Agentic Large Language Model Systems

Year: 2026
Type: article
Abstract: The Agentic large language model (LLM) systems introduce a critical security boundary in which modelgenerated reasoning can trigger consequential external tool actions.This paper presents AgentShield-FW, a hybrid runtime firewall that separates reasoning from execution authority through typed Action Intent Objects, deterministic policy enforcement, classifier-assisted risk assessment, verifier-based route correction, approval and sandbox controls, and tamper-evident auditing.Evaluation combined a 20,000-case deterministic AAF-Bench generator, a leakagecontrolled public corpus, authorization-specific governance tests, adaptive adversarial evaluation, and interactive AgentDojo execution.From 56,155 normalized public records, exact-content deduplication retained 26,129 Clean V2 samples, including a held-out test set of 2,613 cases.On this set, AgentShield-FW achieved 1.77% attack success rate, 98.23% attack blocking, 98.16% F1, 95.29% balanced accuracy, and 97.05% safe completion.A bounded white-box adaptive evaluation produced 0/310 successful malicious trajectories under the specified mutation budget.On the official AgentDojo v1.2.2 subset, injection-task success decreased from 33.33% without protection to 0%, and unsafe canonical tool execution decreased from 92.86% to 0%.However, benign task success decreased from 58.33% to 16.67%, revealing substantial over-blocking.Verifier-specific analysis similarly showed improved malicious containment but measurable benign escalation and route-correctness degradation.The results demonstrate that actionlevel runtime mediation can strongly constrain unsafe agent behaviour while preserving reproducible governance evidence, but they also identify a significant security-utility trade-off that must be addressed before utility-preserving production deployment.
Language: English
Open Access status: hybrid
OpenAlex ID: W7235568065
FWCI:
Cites: 38
Cited by:
Related to: 10