Google's AI Security Agents Uncover 100+ Critical Software Vulnerabilities in 48 Hours

Post Views: 6

Google’s Mandiant has disclosed an AI-powered tool that identifies vulnerabilities in source code, uncovering over 100 high-severity flaws in two days during a live investigation.

The AVDH System

Google’s Mandiant has disclosed the workings of an internal tool named the Agentic Vulnerability Discovery Harness (AVDH), which uses chains of AI agents to hunt for vulnerabilities in source code. The system has identified over 100 verified, high-severity flaws within two days during a live investigation into compromised corporate repositories.

Operational Overview

The AVDH has operated internally at Mandiant for ten months, analyzing tens of millions of lines of code and generating tens of thousands of potential findings. It has identified dozens of exploitable flaws in widely used web extensions and open-source projects, resulting in 12 assigned Common Vulnerabilities and Exposures (CVEs), including CVE-2026-13242 and CVE-2026-55803, with an additional 12 under active disclosure.

Process Breakdown

The AVDH operates as a sequence of specialized AI agents, each passing its output to the next stage, leveraging Google’s Agent Development Kit. The process includes:

Threat Modeling: An agent maps the codebase to determine its type and identifies sections to exclude, such as test directories. A human reviews the threat model before proceeding.
Entry Point Discovery: Agents scan all relevant files to locate points where user input enters the application, such as web routes or inter-process listeners.
Context Enrichment: For each entry point, an agent aggregates scattered code segments, such as permission checks or input sanitizers, that would typically require manual tracing.
Hypothesis Generation: Separate agents analyze potential access-control issues, including unauthorized access, privilege escalation, and cross-site request forgery, as well as risky data flows that could lead to SQL injection, cross-site scripting, command injection, or path traversal.
Hypothesis Validation: Agents operate at elevated temperature settings to expand their reasoning scope, evaluating each hypothesis. A synthesis agent categorizes findings as confirmed, disproven, or rejected. All confirmed results undergo human verification.

Validation and Human Oversight

Mandiant consultants reproduce exploits and execute proof-of-concept code to validate authenticity, discarding findings that fail this test. Mandiant researchers emphasized that network defenders considering similar tools should manually verify results. Traditional automated scanners often generate false positives, as findings may appear plausible but collapse under scrutiny.

Effectiveness and Adaptability

AVDH addresses this by having agents challenge each other’s conclusions and cross-check results against rules crafted by Mandiant consultants, rather than relying on pattern-matching known vulnerabilities. These rules are organized by software domain, language, framework, and vulnerability type, ensuring adaptability across codebases.

Evaluation Methodology

To evaluate its effectiveness, Mandiant developed synthetic, intentionally vulnerable codebases instead of using public datasets, to avoid models recalling pre-trained answers. The researchers highlighted that securing the software development pipeline is a critical challenge in modern enterprise defense.

Conclusion

Manual code reviews cannot match AI’s speed, and conventional scanners frequently miss the breadth of vulnerabilities in contemporary software. However, the tool’s success demonstrates that defenders can regain an edge against adversarial AI by embedding advanced models within expert-designed frameworks. This approach automates the detection of routine vulnerabilities, enabling more efficient threat mitigation.

“Mandiant researchers emphasized that network defenders considering similar tools should manually verify results.”