Google Threat Intelligence Group has revealed that its Agentic Vulnerability Discovery Harness (AVDH) identified more than 100 true-positive critical vulnerabilities in a stolen corporate source-code repository within just two days.
The results demonstrate how agentic artificial intelligence can accelerate vulnerability discovery across large software environments, particularly when organizations face the challenge of assessing compromised repositories and determining which weaknesses could provide attackers with practical exploitation paths.
The development comes as threat actors increasingly adopt AI-assisted capabilities to analyze exposed source code at machine speed.
Google AI Agents Discover 100+ Critical Vulnerabilities
When proprietary repositories are stolen, security teams must rapidly determine whether the code contains exploitable vulnerabilities, identify affected components, and prioritize remediation.
Google said AVDH was developed to help defenders move from broad code visibility toward prioritized findings that can subsequently be validated by human security experts.
AVDH combines large language models with a structured orchestration layer based on Google’s Agent Development Kit. Instead of depending on a single AI prompt or traditional signature-based scanning, the framework uses multiple specialized agents operating through a sequential vulnerability-discovery process.
This architecture allows different agents to examine application behavior, entry points, access controls, data flows, and potential exploitation paths before findings are consolidated for validation.
Google said Mandiant teams have used the framework during the past 10 months across environments containing tens of millions of lines of source code.
The system has executed thousands of analysis pipelines and produced tens of thousands of potential findings. The research has contributed to 12 assigned CVEs affecting widely deployed web extensions and open-source software, including Drupal vulnerabilities CVE-2026-13242 and CVE-2026-55803. Additional vulnerabilities remain under disclosure.
The discovery process begins with threat modeling. An Explorer agent determines the application’s purpose, technology stack, documentation, relevant directories, and areas that should be excluded from analysis.
Specialist agents then inspect security-sensitive functionality such as authentication, authorization, routing, and application-specific components.
Their findings are combined into a threat model that is reviewed by a human consultant before deeper automated analysis begins.
Discovery agents subsequently analyze relevant source-code files in parallel to identify application entry points. These can include HTTP routes, inter-process communication listeners, and other externally accessible interfaces.
The system then identifies user-controlled inputs associated with those entry points, creating the foundation for deeper access-control and data-flow analysis.
Enrichment agents gather additional context that may be distributed across different functions and files. This capability is important because authorization checks, sanitization routines, routing restrictions, and security controls are frequently implemented in separate parts of an application.
By connecting these components, AVDH can determine whether sensitive functionality is exposed to unauthorized users or whether attacker-controlled data can reach dangerous execution points.
The framework uses specialized agents for access-control and data-flow analysis. Access-control analysis can identify weaknesses involving missing authorization checks, privilege escalation, and cross-site request forgery.
Data-flow analysis traces untrusted input through application logic, transformations, databases, and execution paths to identify potential SQL injection, cross-site scripting, command injection, and path-traversal vulnerabilities.
Potential findings are independently validated before a synthesis agent determines whether the available evidence supports or rejects each vulnerability hypothesis. Google also applies confidence filters and varied model settings to improve coverage while limiting false positives.
Human expertise remains central to the final validation stage. Mandiant consultants reproduce suspected attack paths and execute proof-of-concept code where appropriate to determine whether vulnerabilities can actually be exploited and whether compensating controls prevent successful attacks. Findings that fail validation are discarded.
Google’s research highlights a growing role for agentic AI in defensive cybersecurity. Rather than replacing experienced security professionals.
Automated source-code analysis can act as a force multiplier, allowing defenders to process enormous codebases faster and focus human expertise on complex exploit chains, business-logic weaknesses, and real-world attack scenarios that still require careful judgment.
No Comment! Be the first one.