Staying Ahead of Adversarial AI Through Agentic Source Code Review
Mandiant introduced the Agentic Vulnerability Discovery Harness (AVDH), a multi-agent orchestration framework built on Google's Agent Development Kit (ADK) for rapid, high-precision source code vulnerability analysis. AVDH uses a sequential, waterfall-style pipeline of specialized agents—starting with an Explorer agent for codebase reconnaissance, followed by Specialist Explorers, and a Threat Model Synthesis agent—to systematically identify and prioritize vulnerabilities. In 10 months of operat
Analysis
TL;DR
- Mandiant introduced the Agentic Vulnerability Discovery Harness (AVDH), a multi-agent orchestration framework built on Google's Agent Development Kit (ADK) for rapid, high-precision source code vulnerability analysis.
- AVDH uses a sequential, waterfall-style pipeline of specialized agents—starting with an Explorer agent for codebase reconnaissance, followed by Specialist Explorers, and a Threat Model Synthesis agent—to systematically identify and prioritize vulnerabilities.
- In 10 months of operation, AVDH discovered over 100 true-positive critical vulnerabilities in two days during an incident response investigation involving stolen corporate repositories, and has produced 12 assigned CVEs with a dozen more in active disclosure.
- The framework combines LLM-powered code analysis with rich environmental context (SBOMs, architecture docs, threat intelligence) and human expert validation gates to mitigate LLM non-determinism and improve accuracy.
- AVDH is designed to complement existing scanning tools like CodeMender, forming a two-layered defense strategy that accelerates both broad-scale vulnerability discovery and targeted adversary simulation engagements.
Why It Matters
Mandiant's AVDH represents a significant step forward in applying agentic AI to real-world cybersecurity operations, demonstrating that structured multi-agent pipelines can outperform both manual review and traditional static analysis at scale. For AI practitioners and security teams, it provides a concrete, shareable architecture for building LLM-driven vulnerability discovery systems that balance automation with human oversight. The results—100+ critical vulnerabilities found in two days and 12 CVEs—prove that agentic approaches can meaningfully close the gap between defenders and adversaries leveraging AI for offensive purposes.
Technical Details
- Framework & Orchestration: AVDH is built on Google's Agent Development Kit (ADK), implementing sequential multi-agent orchestration patterns. The architecture aligns with capabilities in Google Antigravity for centralized workflow management. Agents are chained in a deterministic pipeline where each phase completes before the next begins, reducing the non-deterministic risks inherent in LLM outputs.
- Pipeline Architecture: The system follows a waterfall-style sequential methodology. It begins with an Explorer agent that identifies the codebase's core purpose, software domain, and directories to exclude, then dispatches Specialist Explorer subagents focused on authentication, authorization, routing, and other domain-specific areas. Findings are aggregated by a Threat Model Synthesis agent into a cohesive threat model, which undergoes human consultant verification before proceeding.
- Context Integration: AVDH ingests rich environmental inputs including asset inventories, software bills of materials (SBOMs), architecture documentation, and threat intelligence. This contextual data is distilled into a knowledge base that enables agents to dynamically select relevant skills, language rules, and vulnerability patterns for deep, targeted analysis.
- Validation & Human-in-the-Loop: The pipeline incorporates skeptical validation steps and human expert approval gates—particularly after threat model synthesis—to ensure accuracy before deeper analysis. This hybrid approach allows AI to handle scale while humans focus on high-impact validation and complex exploit chain reasoning.
- Benchmarks & Results: Over 10 months, AVDH analyzed environments spanning tens of millions of lines of code, executed thousands of pipelines, and generated tens of thousands of findings. It uncovered assignable flaws in widely used web extensions and open-source projects, resulting in 12 assigned CVEs (including CVE-2026-13242 and CVE-2026-55803) with a dozen more in active disclosure. During incident response, it identified over 100 true-positive critical vulnerabilities in two days, and in adversary simulations, it quickly discovered an RCE vulnerability enabling initial access.
Industry Insight
- The AVDH architecture demonstrates that the most effective agentic security systems are not fully autonomous but rather structured pipelines with human validation gates—organizations should prioritize deterministic orchestration and expert-in-the-loop designs over fully autonomous AI agents for critical security tasks.
- As adversarial AI misuse accelerates, the combination of agentic source code review with continuous scanning tools (like CodeMender) creates a scalable two-layered defense; security teams should evaluate similar hybrid approaches that pair rapid agentic analysis with ongoing automated scanning for comprehensive coverage.
- The public sharing of AVDH's architecture by Mandiant signals a broader industry shift toward open, reproducible agentic security frameworks—organizations should invest in building internal agentic capabilities tailored to their specific codebases and threat landscapes rather than relying solely on off-the-shelf scanning tools.
Disclaimer: The above content is generated by AI and is for reference only.