Beyond SAST and DAST: A Unified Security Testing Architecture for Autonomous Coding Agents

Autonomous coding agents-AI systems that independently write, review, commit, and deploy software-have introduced a class of security vulnerabilities that existing application security testing methodologies cannot detect. Static Application Security Testing (SAST) and Dynamic Application Security Testing (DAST), the twin pillars of modern secure development lifecycles, were designed under a foundational assumption: human developers author code, and the development pipeline consists of discrete, human-initiated events. Agentic AI systems invalidate both assumptions. This paper argues that SAST and DAST must be fundamentally reinvented, not incrementally extended, to address the threat landscape of agentic code generation. We demonstrate this gap through analysis of a representative supply chain attack vector-the adversarially poisoned AI-suggested dependency-which evades all major SAST and DAST controls by exploiting the trust model of the agent rather than the syntax or runtime behavior of the generated code. We present three contributions: (1) the Agentic Vulnerability Taxonomy (AVT), classifying vulnerability classes unique to or amplified by agentic coding pipelines across five dimensions; (2) the Agentic Risk Scoring Model (ARSM), a quantitative framework extending CVSS 3.1 with four agent-specific dimensions; and (3) the Unified Agentic Security Testing (UAST) architecture, a redesigned security testing pipeline integrating static, dynamic, behavioral, and provenance-aware analysis into a continuous, agent-aware feedback loop. These contributions are aligned with NIST AI RMF [1], ISO/IEC 42001 [2], and OWASP guidance on LLM application security [3].