Threat Intelligence Stays Ahead of Adversarial AI With Agentic Source Code Review, Google Says

KUALA LUMPUR, 19 AUGUST 2026 —

Google Unveils AI-Powered Code Review Tool to Counter Adversarial Threats

The framework has already demonstrated significant impact, including the discovery of over 100 critical vulnerabilities in just two days during a recent incident response investigation involving stolen corporate repositories.

This rapid analysis—achieved in a fraction of the time required for manual review—has enabled Google’s Mandiant team to scale vulnerability discovery across tens of millions of lines of code, generating tens of thousands of findings and uncovering dozens of assignable flaws in widely used web extensions and open-source projects. These efforts have resulted in 12 assigned CVEs, including CVE-2026-13242 and CVE-2026-55803, with additional disclosures currently in progress.

AVDH operates as a two-layered defense strategy, complementing ongoing scanning tools like CodeMender. By structuring the analysis process with skeptical validation steps and injecting domain-specific human expertise, Google claims the framework helps defenders "beat adversaries to the punch" in high-stakes cybersecurity engagements.

The tool has also proven critical in adversary simulation exercises, where it quickly identified a remote code execution (RCE) vulnerability in a client’s web application, enabling initial access—a scenario that would typically require extensive manual effort.

How AVDH Works: A Structured, Agentic Approach At its core, AVDH leverages the Google Agent Development Kit (ADK), an LLM framework that implements agent orchestration patterns to mitigate unpredictability in large language models (LLMs). The system chains specialized agents in a sequential pipeline—akin to a waterfall software development model—where each phase must complete before the next begins. This structured approach ensures a prioritized, risk-rated list of findings ready for human expert review.

The process begins with an Explorer agent, which identifies the target codebase’s purpose, software domain (e.g., web or desktop application), and flags irrelevant directories (such as unit tests). Specialist Explorer subagents then analyze domain-specific areas like authentication, authorization, and routing, feeding their findings into a Threat Model Synthesis agent. This agent aggregates the data into a cohesive threat model, presented to consultants for verification before proceeding.

The visual threat model (as shown in Figure 3 of Google’s report) maps exposed application components and their interconnections, providing critical context for subsequent analysis.

With the threat model established, parallelized Discovery agents deploy the lightweight Gemini Flash Lite model to scan in-scope files, extracting application entry points such as HTTP routes and inter-process communication (IPC) listeners. These agents simultaneously isolate user input sources nested within these entry points. Each entry point is then assigned to an Enrichment agent, which navigates the codebase to aggregate contextually relevant data—such as sanitizers, permissions, and routing conditions—often buried deep within nested function calls.

Control Flow and Data Flow: The Heart of Vulnerability Detection

Effective code analysis hinges on two primary properties: control flow (the execution order of tasks) and data flow (how information moves and transforms). AVDH delegates these tasks to two specialized agents:

1. Access Control Agent: Evaluates protections around target entry points to determine accessibility. It validates security assumptions, confirming whether privileged functionality is restricted or inadvertently exposed to unauthorized users. This analysis exposes flaws such as missing authorization, privilege escalation, and cross-site request forgery (CSRF).

2. Data Flow Analysis Agent: Traces user input from the initial entry point throughout the application, identifying potential injection points, insecure deserialization, and other data-driven vulnerabilities.

To manage the volume of hypotheses generated, a Confidence Filter—configured by consultants—keeps the creative process in check. This ensures that only high-probability findings proceed to validation, reducing noise and improving efficiency.

Regional and Industry Implications The introduction of AVDH comes at a time when AI-driven cyber threats are escalating globally, with adversaries increasingly weaponizing AI to automate attacks. For Malaysia, where digital transformation in sectors like finance, healthcare, and government is accelerating, the risks are particularly acute. The country’s Critical National Information Infrastructure (CNII)—encompassing utilities, telecommunications, and financial services—remains a prime target for sophisticated cyber threats.

Google’s framework could serve as a force multiplier for Malaysian cybersecurity teams, particularly those grappling with open-source vulnerabilities and third-party software risks. The 12 assigned CVEs uncovered by AVDH highlight the prevalence of exploitable flaws in widely used tools, a concern for organizations that rely on open-source components.

In Malaysia, where government-linked companies (GLCs) and financial institutions are increasingly adopting open-source solutions, proactive vulnerability management is critical to preventing breaches that could disrupt services or lead to data theft.

Industry stakeholders in Malaysia have welcomed the development. Dr. Solahuddin Shamsuddin, Chief Technology Officer at CyberSecurity Malaysia, noted that tools like AVDH align with the country’s National Cyber Security Strategy 2023-2028, which emphasizes AI-driven threat detection and response. "The ability to rapidly analyze millions of lines of code and identify critical vulnerabilities in near real-time is a game-changer," he said.

"For Malaysian organizations, especially those in CNII sectors, this could reduce the window of exposure to cyber threats."

Challenges and Considerations While AVDH represents a leap in automated vulnerability discovery, Google acknowledges limitations. The framework’s effectiveness depends on high-quality contextual inputs, such as asset inventories, software bills of materials (SBOMs), architecture documentation, and threat intelligence. Without these, the risk of false positives or missed vulnerabilities increases. The tool’s reliance on LLMs introduces non-deterministic behavior, which Google mitigates through structured orchestration and human oversight.

For Malaysian enterprises, adoption may face hurdles such as integration with existing security tools, training requirements for cybersecurity teams, and cost considerations. However, Google’s decision to share the architecture publicly suggests an intent to foster broader industry adoption, potentially lowering barriers for organizations with limited resources.

Looking Ahead: The Future of AI in Cybersecurity

Google’s AVDH is part of a broader trend toward agentic cybersecurity, where AI systems autonomously execute complex workflows to augment human analysts. The company’s Antigravity platform, which provides a centralized workspace for managing agentic workflows, further underscores this shift. As AI tools become more sophisticated, the cybersecurity arms race will intensify, with defenders leveraging automation to stay ahead of adversaries.

For Malaysia, the adoption of such frameworks could enhance resilience against AI-powered threats, particularly as the country positions itself as a regional digital hub. However, success will depend on collaboration between government, industry, and academia to ensure that cybersecurity capabilities evolve in tandem with technological advancements.

As Google’s Threat Intelligence Group continues to refine AVDH, the framework’s real-world impact—measured in reduced incident response times, lower breach risks, and faster patching cycles—will determine its long-term viability. For now, the tool stands as a testament to the potential of AI-human collaboration in cybersecurity, offering a glimpse into the future of proactive threat hunting.

Details not yet available on whether Malaysian organizations have begun piloting AVDH or if Google plans to offer localized support for the framework.

Related: Google