Google Mandiant AI Agents Find Over 100 Critical Vulnerabilities in Source Code Within Two Days

Google’s Threat Intelligence Group has announced that its Agentic Vulnerability Discovery Harness (AVDH) identified over 100 critical true-positive vulnerabilities in stolen corporate source code repositories within just two days.

This result highlights how agentic AI can significantly accelerate vulnerability discovery during incident response, red teaming, penetration testing, and proactive secure code reviews, especially when adversaries already have access to proprietary code and can leverage AI-assisted analysis at machine speed.

Google Mandiant Find Critical Vulnerabilities

According to Google, the AVDH combines specialized large language model (LLM) agents with deterministic orchestration and expert-defined security methodologies.

Rather than functioning as a standalone code scanner, the system organizes agents into a sequential pipeline that produces prioritized, risk-rated findings for human validation.

Google has been using this framework for 10 months in environments containing tens of millions of lines of code. It has executed thousands of analysis pipelines and generated tens of thousands of findings, resulting in the assignment of 12 Common Vulnerabilities and Exposures (CVEs) in web extensions and open-source projects.

Notable examples include CVE-2026-13242 and CVE-2026-55803, with additional flaws reportedly undergoing coordinated disclosure.

The AVDH begins with an Explorer agent that profiles a target codebase, assesses its purpose, reviews documentation, and determines which directories to exclude, such as unit test folders.

Specialist Explorer agents then focus on specific areas, including authentication, authorization, routing, and software-domain-specific attack surfaces.

A Threat Model Synthesis agent consolidates the results from these agents. This stage generates both textual and visual threat models that identify exposed components, trust boundaries, likely attack vectors, and application relationships. Importantly, a human consultant must approve the threat model before deeper analysis commences.

After threat modeling, parallel Discovery agents scan the in-scope files to identify entry points such as HTTP routes, IPC listeners, and other externally reachable interfaces. These agents also isolate user-controlled input sources that could access sensitive functions.

An Enrichment agent expands the analysis beyond isolated entry points by tracing relevant code across files and nested calls, collecting context related to sanitization, permissions, routing logic, and storage boundaries.

This addresses a common limitation of both manual review and conventional scanning: exploitable conditions often arise across distributed code paths.

The harness assigns control-flow analysis to Access Control agents and data-flow tracking to Data Flow Analysis agents. The Access Control agents search for missing or broken authorization checks, privilege escalation pathways, and CSRF exposure.

In contrast, Data Flow Analysis agents trace user input to dangerous sinks, highlighting potential SQL injection, cross-site scripting, command injection, and path traversal flaws.

Google’s workflow intentionally separates broad hypothesis generation from strict validation. Multiple Validation agents assess potential vulnerabilities using higher-temperature configurations to explore various reasoning paths.

A Validation Synthesis agent reviews the gathered evidence and classifies each issue as confirmed, disproven, or rejected.

Confirmed findings are deduplicated and risk-rated before being passed to human experts. These consultants dynamically test exploitation paths and proof-of-concept code to eliminate false positives and identify compensating controls that AI may not recognize.

Google emphasizes that AVDH is not meant to replace security researchers. Instead, its effectiveness depends on the “distilled knowledge” derived from consultants’ experiences, including reusable language, frameworks, and vulnerability-specific rules.

The company advises organizations looking to adopt similar systems to establish strong human validation processes and proprietary benchmark environments. Public vulnerability datasets can be compromised by LLM training data, complicating the measurement of genuine reasoning performance.

The results reported by AVDH underscore an increasing necessity: organizations require AI-assisted source code analysis that can keep pace with the speed of adversarial AI, while still retaining human oversight for decisions about exploitability, disclosure, and remediation.

Prevent incidents due to slow investigations. Power your Tier 1 with threat intelligence from 15K SOCs: Integrate TI Lookup in your SOC