Staying Ahead of Adversarial AI Through Agentic Source Code Review

Adversarial misuse of AI has heightened risks of data theft and extortion, particularly when proprietary source code is exposed. Google Cloud Threat Intelligence’s new Agentic Vulnerability Discovery Harness (AVDH) addresses this by integrating AI models with structured, human-driven orchestration to accelerate vulnerability discovery. The framework analyzes code during proactive reviews, penetration tests, and incident responses, enabling defenders to identify exploit paths more efficiently than manual methods. AVDH has been deployed for 10 months, analyzing tens of millions of lines of code and uncovering over 100 critical vulnerabilities in a recent incident response case within two days.

The harness operates as a two-layered defense alongside tools like CodeMender, combining multi-agent orchestration with frontline expertise to prioritize findings for human review. It has identified dozens of assignable flaws in widely used projects, resulting in 12 assigned CVEs, including CVE-2026-13242 and CVE-2026-55803. AVDH’s deterministic pipeline, built using the Google Agent Development Kit (ADK), mitigates unpredictability in large language models (LLMs) and improves code analysis effectiveness. The framework chains specialized agents in a sequential pipeline, mirroring waterfall development, to produce risk-rated findings for expert evaluation.

A critical first step in AVDH’s process is establishing a threat model for the target codebase, distinguishing between accessible and restricted code while identifying attack vectors. The pipeline begins with an Explorer agent that determines the software domain, reviews documentation, and flags irrelevant directories. Specialist Explorer subagents then focus on areas like authentication and routing, feeding their findings into a Threat Model Synthesis agent that generates both textual and visual threat models for consultant verification before proceeding.

Once the threat model is validated, Discovery agents use the lightweight Gemini Flash Lite model to identify application entry points such as HTTP routes and IPC listeners. Enrichment agents then aggregate contextually relevant code to assess whether further analysis by Access Control or Data Flow Analysis agents is required. The Access Control agent evaluates permissions and routing conditions, while the Data Flow Analysis agent traces information movement, ensuring comprehensive coverage of control and data flow properties essential for effective code security analysis.