Google AVDH: Agentic Source Code Review Against Adversarial AI

Adversarial misuse of AI is compressing the window defenders have to identify and patch vulnerabilities before attackers weaponize them. Google's Agentic Vulnerability Discovery Harness (AVDH) reframes source code review as an agent-orchestrated pipeline guided by human expertise.

When proprietary source code is exposed, defenders must scramble while attackers deploy machine-speed AI tools against them. Mandiant and Google Cloud have introduced AVDH, an internal harness that structures LLM-driven analysis, enforces skeptical validation, and injects domain expertise directly into the workflow. The result: faster, higher-precision vulnerability discovery across proactive reviews, penetration tests, red team operations, and incident response engagements.

What was announced

Google's Threat Intelligence team disclosed the architecture of the Agentic Vulnerability Discovery Harness (AVDH), used internally by Mandiant for the past 10 months. During a recent incident response investigation involving stolen corporate repositories, AVDH discovered more than 100 true-positive critical vulnerabilities in just two days — a fraction of the time required for manual review.

Across engagements, the harness has analyzed environments spanning tens of millions of lines of code, executed thousands of pipelines, and generated tens of thousands of findings. Assignable flaws in widely used web extensions and open-source projects have resulted in 12 assigned CVEs, including CVE-2026-13242 and CVE-2026-55803, with an additional dozen currently in active disclosure. AVDH is designed to be used alongside CodeMender's ongoing scanning for a two-layered defense strategy.

Why this matters

For CIOs, CISOs, IT directors, and procurement leaders, AVDH signals a maturity shift in AI-assisted security. Manual source code review cannot keep pace with adversarial AI, and traditional scanning engines miss many modern vulnerability classes. By embedding frontier models within an expert-defined harness, defenders automate the discovery of routine vulnerabilities and free specialists to focus on complex exploit chains.

AI is most effective when deployed as a practical multiplier for human expertise, transforming source code visibility into a scalable defense

Technical details

• Framework: Built on the Google Agent Development Kit (ADK), aligned with agentic orchestration in Google Antigravity.
• Sequential pipeline: Explorer agents perform reconnaissance and produce a threat model reviewed by a human before analysis continues.
• Entry point discovery: Parallelized Discovery agents use Gemini Flash Lite to extract HTTP routes, IPC listeners, and other attack vectors at scale.
• Context enrichment: Dedicated Enrichment agents aggregate distributed sanitizers, permissions, and routing conditions across nested calls.
• Hypothesis generation: Access Control and Data Flow Analysis agents brainstorm potential flaws — SQL injection, XSS, command injection, path traversal, CSRF, privilege escalation.
• Validation: Multiple high-temperature Validation agents feed a Synthesis agent that confirms, disproves, or rejects each hypothesis.
• Human-in-the-loop: Consultants replicate exploitation with POC code before formal disclosure.
• Distilled knowledge: A modular rules hierarchy (domain, language, framework, vulnerability) injects Mandiant expertise into every stage.
• Benchmarking: Proprietary synthetic codebases avoid training data contamination and measure precision, false positives, and duplicates.

Softprom and Google

Softprom is the official partner of Google. Our team helps enterprises evaluate, deploy, and operationalize Google Cloud security capabilities — including AI-driven threat defense, secure development pipelines, and vulnerability management aligned with modern SOC workflows.

This content was prepared as part of the Softprom DistriFlow project — an automated system for monitoring and adapting vendor news. Original source: original article.