ZT-ICAS: A Zero-Trust Integrity-Constrained Framework for Agentic Vulnerability Scanning

Software vulnerability discovery is currently dominated by static application security testing (SAST) and heuristic rule-based scanners. While scalable, these approaches suffer from high false-positive rates, weak exploitability assessment, and limited adaptability to evolving codebases. Recent advances in Agentic AI offer a paradigm shift, enabling adaptive scanning through autonomous tool orchestration and multi-step reasoning. However, embedding agentic reasoning into security workflows fundamentally expands the attack surface, as these systems must ingest adversarially controllable inputs such as source code, inline comments, and dependency metadata to function. This exposure allows attackers to exploit the scanner via prompt injection and reasoning manipulation. This paper presents a secure-by-design framework for agentic vulnerability scanning that explicitly treats the analysis system as a security-critical target. We propose a Zero-Trust architecture enforcing semantic input neutralization, ephemeral context sharding, and capability-bounded tool invocation to preserve analysis integrity. We empirically evaluate this framework against codebases augmented with adversarial comments and poisoned metadata. Results demonstrate that while unconstrained agents are susceptible to systematic manipulation, our constraints significantly reduce analysis compromise with manageable performance overhead.