Google's AI Agents Discover Over 100 Critical Software Flaws in Two Days

Google's Mandiant has revealed the inner workings of its Agentic Vulnerability Discovery Harness (AVDH), an internal tool leveraging chains of specialized AI agents to proactively hunt for software vulnerabilities. In a recent, real-world investigation targeting stolen corporate code repositories, AVDH demonstrated remarkable efficiency by identifying more than 100 high-severity flaws within a mere two-day period. This breakthrough highlights the accelerating pace at which AI can be employed in cybersecurity defense.

AVDH has been operational within Mandiant for the past ten months, during which it has meticulously scanned tens of millions of lines of code. This extensive analysis has yielded tens of thousands of potential findings, according to a recent publication by the Google Threat Intelligence Group. The tool has already been credited with uncovering dozens of assignable vulnerabilities in widely adopted web extensions and open-source projects, leading to the assignment of 12 CVEs, including CVE-2026-13242 and CVE-2026-55803, with an additional dozen currently undergoing active disclosure.

The AVDH pipeline operates as a sophisticated sequence of specialized AI agents, each building upon the output of the preceding one. This framework, constructed using Google's Agent Development Kit, begins with a threat modeling phase where an agent maps the codebase, identifies the software type, and flags sections to be excluded, such as test directories. Crucially, a human expert reviews this threat model before the automated process continues.

Following threat modeling, the pipeline moves to entry point discovery, where agents systematically scan all relevant files to pinpoint locations where user input enters the application. This includes web routes and inter-process listeners. Subsequently, the context enrichment stage gathers scattered but relevant code snippets associated with each entry point, such as permission checks and input sanitizers, which would typically require significant manual effort for a human reviewer to locate.

The core of AVDH's vulnerability detection lies in its hypothesis generation and validation stages. Separate agents are tasked with identifying potential access-control issues, including missing authorization, privilege escalation, and cross-site request forgery, as well as dangerous data flows that could lead to common exploits like SQL injection, cross-site scripting, command injection, and path traversal. Multiple agents, operating with high "temperature" settings to encourage broader reasoning, then weigh in on these hypotheses.

A synthesis agent consolidates the agents' assessments, categorizing each hypothesis as confirmed, disproven, or rejected. However, even confirmed findings are not finalized until they undergo rigorous human scrutiny. Mandiant consultants meticulously reproduce the exploit and execute proof-of-concept code to verify the genuine existence of the flaw and ensure no overlooked controls mitigate it. Any finding that fails this final human validation is discarded.

Mandiant emphasizes that AVDH was specifically designed to combat the issue of false positives, a common problem with traditional automated code scanners. By enabling agents to challenge each other's conclusions and cross-reference findings against rules established by their own consultants, AVDH aims for higher accuracy than tools that merely flag code patterns resembling known bugs.

"Securing the software development pipeline has emerged as a defining challenge in modern enterprise defense," the researchers stated. "To match these emerging threats, securing the code pipeline must be a critical component of a modern defense strategy. Manual source code review can’t keep pace with AI, and traditional scanning engines consistently miss the broad spectrum of vulnerabilities hidden in modern software." The success of AVDH, they conclude, demonstrates that defenders can indeed regain the advantage against adversarial AI by embedding advanced models within expert-defined harnesses to automate the discovery of routine vulnerabilities.