Google's Mandiant security division has revealed an internal AI-driven tool designed to identify software vulnerabilities, which successfully uncovered over 100 verified, high-severity flaws in just two days during a live investigation of stolen corporate code repositories. The tool, named the Agentic Vulnerability Discovery Harness (AVDH), has been operational within Mandiant for ten months, during which it has processed tens of millions of lines of code and generated tens of thousands of potential findings.
Mandiant researchers Alex Tselevich and Michael Maturi stated that the AVDH has led to the discovery of numerous assignable flaws in widely used web extensions and open-source projects. This effort has resulted in 12 assigned Common Vulnerabilities and Exposures (CVEs), including CVE-2026-13242 and CVE-2026-55803, with an additional dozen currently undergoing active disclosure.
The AVDH operates as a sequence of specialized AI agents, built on Google's Agent Development Kit, with each agent passing its output to the next. The process begins with a threat modeling agent that maps the codebase, identifies the software type, and designates areas to be excluded, such as test directories. This initial threat model is then reviewed by a human operator.
Following this, an entry point discovery agent scans all in-scope files to locate points where user input enters the application, ranging from web routes to inter-process listeners. A context enrichment agent then gathers relevant, scattered code—such as permission checks and input sanitizers—for each identified entry point, which would otherwise require manual collection.
Hypothesis generation agents then search for various types of vulnerabilities. Some agents focus on access-control issues, including missing authorization, privilege escalation, and cross-site request forgery. Others concentrate on dangerous data flows that could lead to SQL injection, cross-site scripting, command injection, and path traversal.
The subsequent stage involves hypothesis validation, where multiple agents, intentionally run with high "temperature" settings to encourage diverse reasoning, evaluate each generated hypothesis. A synthesis agent then categorizes each finding as confirmed, disproven, or rejected. Crucially, every confirmed finding undergoes human validation, where Mandiant consultants reproduce the exploit and execute proof-of-concept code to ensure the flaw is genuine and not mitigated by any overlooked controls. Findings that fail this human verification are discarded.
Mandiant emphasized that the AVDH was specifically designed to address the problem of "noise" or false positives common in automated code scanners. It achieves this by having agents challenge each other's conclusions and verify findings against rules established by Mandiant's consultants, rather than merely flagging code patterns that resemble known bugs. These rules are organized by software domain, then further categorized by language, framework, and vulnerability type, ensuring reusability across different codebases.
To evaluate its own performance, Mandiant developed a set of synthetic, intentionally vulnerable codebases. This approach was chosen over relying on public vulnerability datasets, due to concerns that current AI models might have been trained on such datasets and could be recalling answers rather than genuinely reasoning through them.
The researchers highlighted that securing the software development pipeline is a critical challenge in modern enterprise defense, and manual source code review cannot keep pace with the speed of AI. They asserted that traditional scanning engines frequently miss a broad spectrum of vulnerabilities in modern software. Mandiant believes that the success of their harness demonstrates that defenders can regain an advantage against adversarial AI by embedding frontier models within an expert-defined framework to automate the discovery of routine vulnerabilities. They encourage other network defenders considering similar vulnerability discovery harnesses to manually validate all findings.
