The Agentic Vulnerability Discovery Harness, or AVDH, combines multiple specialised AI agents with structured orchestration and human security expertise. Mandiant has deployed the system for about 10 months across incident-response investigations, penetration tests, red-team exercises and proactive source-code reviews.

The result involving more than 100 critical vulnerabilities emerged during an investigation into corporate repositories stolen by attackers. The findings were classified as true positives rather than unverified alerts, an important distinction because automated security scanners can produce large volumes of findings that ultimately prove harmless. Human specialists remained involved in confirming exploitability.

AVDH has been used to examine software environments containing tens of millions of lines of code. Thousands of analysis pipelines have produced tens of thousands of potential findings, while deeper investigation has exposed dozens of flaws capable of qualifying for formal vulnerability identifiers. Twelve Common Vulnerabilities and Exposures, or CVEs, have already been assigned, while roughly another dozen flaws are moving through coordinated disclosure processes.

Among the assigned vulnerabilities are CVE-2026-13242 and CVE-2026-55803. The technology has also identified weaknesses in widely used web extensions and open-source projects, widening its potential relevance beyond investigations of proprietary enterprise software.

Rather than asking a single large language model to inspect an entire application, the system divides vulnerability discovery into a series of specialised tasks. An initial agent examines the purpose and structure of the codebase, studies available documentation and identifies directories that should be excluded from analysis, such as testing components. Other agents focus on areas including authentication, authorisation and routing.

Their findings are consolidated into a threat model that must be reviewed by a human consultant before the automated investigation proceeds. Discovery agents then analyse files for potential entry points, including HTTP routes, inter-process communication listeners and other locations where user-controlled data can enter an application.

Additional agents collect surrounding context that could determine whether apparent weaknesses are genuinely exploitable. This includes permission checks, sanitisation functions, routing conditions and nested function calls. The approach is intended to address a long-standing limitation of conventional static-analysis tools, which can identify suspicious coding patterns without always understanding whether vulnerable code can actually be reached by an attacker.

The system separately evaluates control flow and data flow. Control-flow analysis can identify weaknesses such as missing authorisation checks, privilege escalation opportunities and cross-site request forgery. Data-flow agents trace untrusted input through an application to determine whether it can reach sensitive operations capable of enabling SQL injection, cross-site scripting, command injection or path traversal.

Potential vulnerabilities then pass through several validation agents before being confirmed, rejected or marked for further examination. Verified findings are deduplicated and assigned risk levels. Security specialists subsequently attempt to reproduce the weakness and validate proof-of-concept exploits, preserving human oversight at the stage where an automated hypothesis becomes an actionable security finding.

AVDH has also demonstrated value in offensive-security simulations. During one customer engagement, analysis of a web application uncovered a remote-code-execution vulnerability that provided a route for initial access, showing how the same technology can help red teams locate weaknesses that attackers might exploit.

The development reflects a wider shift towards multi-agent security systems rather than reliance on individual AI models. Research published this year has shown that the design of the surrounding agent harness — including roles, tools, communication and verification procedures — can substantially influence vulnerability-discovery performance even when the underlying language model remains unchanged.

Google is positioning AVDH as a point-in-time tool for intensive investigations rather than continuous software monitoring. Its CodeMender technology is designed to complement that approach by scanning code, validating exploitable weaknesses and generating tested patches within development workflows.

The Agentic Vulnerability Discovery Harness, or AVDH, combines multiple specialised AI agents with structured orchestration and human security expertise. Mandiant has deployed the system for about 10 months across incident-response investigations, penetration tests, red-team exercises and proactive source-code reviews.

The result involving more than 100 critical vulnerabilities emerged during an investigation into corporate repositories stolen by attackers. The findings were classified as true positives rather than unverified alerts, an important distinction because automated security scanners can produce large volumes of findings that ultimately prove harmless. Human specialists remained involved in confirming exploitability.

AVDH has been used to examine software environments containing tens of millions of lines of code. Thousands of analysis pipelines have produced tens of thousands of potential findings, while deeper investigation has exposed dozens of flaws capable of qualifying for formal vulnerability identifiers. Twelve Common Vulnerabilities and Exposures, or CVEs, have already been assigned, while roughly another dozen flaws are moving through coordinated disclosure processes.

Among the assigned vulnerabilities are CVE-2026-13242 and CVE-2026-55803. The technology has also identified weaknesses in widely used web extensions and open-source projects, widening its potential relevance beyond investigations of proprietary enterprise software.

Rather than asking a single large language model to inspect an entire application, the system divides vulnerability discovery into a series of specialised tasks. An initial agent examines the purpose and structure of the codebase, studies available documentation and identifies directories that should be excluded from analysis, such as testing components. Other agents focus on areas including authentication, authorisation and routing.

Their findings are consolidated into a threat model that must be reviewed by a human consultant before the automated investigation proceeds. Discovery agents then analyse files for potential entry points, including HTTP routes, inter-process communication listeners and other locations where user-controlled data can enter an application.

Additional agents collect surrounding context that could determine whether apparent weaknesses are genuinely exploitable. This includes permission checks, sanitisation functions, routing conditions and nested function calls. The approach is intended to address a long-standing limitation of conventional static-analysis tools, which can identify suspicious coding patterns without always understanding whether vulnerable code can actually be reached by an attacker.

The system separately evaluates control flow and data flow. Control-flow analysis can identify weaknesses such as missing authorisation checks, privilege escalation opportunities and cross-site request forgery. Data-flow agents trace untrusted input through an application to determine whether it can reach sensitive operations capable of enabling SQL injection, cross-site scripting, command injection or path traversal.

Potential vulnerabilities then pass through several validation agents before being confirmed, rejected or marked for further examination. Verified findings are deduplicated and assigned risk levels. Security specialists subsequently attempt to reproduce the weakness and validate proof-of-concept exploits, preserving human oversight at the stage where an automated hypothesis becomes an actionable security finding.

AVDH has also demonstrated value in offensive-security simulations. During one customer engagement, analysis of a web application uncovered a remote-code-execution vulnerability that provided a route for initial access, showing how the same technology can help red teams locate weaknesses that attackers might exploit.

The development reflects a wider shift towards multi-agent security systems rather than reliance on individual AI models. Research published this year has shown that the design of the surrounding agent harness — including roles, tools, communication and verification procedures — can substantially influence vulnerability-discovery performance even when the underlying language model remains unchanged.

Google is positioning AVDH as a point-in-time tool for intensive investigations rather than continuous software monitoring. Its CodeMender technology is designed to complement that approach by scanning code, validating exploitable weaknesses and generating tested patches within development workflows.

The Agentic Vulnerability Discovery Harness, or AVDH, combines multiple specialised AI agents with structured orchestration and human security expertise. Mandiant has deployed the system for about 10 months across incident-response investigations, penetration tests, red-team exercises and proactive source-code reviews.

The result involving more than 100 critical vulnerabilities emerged during an investigation into corporate repositories stolen by attackers. The findings were classified as true positives rather than unverified alerts, an important distinction because automated security scanners can produce large volumes of findings that ultimately prove harmless. Human specialists remained involved in confirming exploitability.

AVDH has been used to examine software environments containing tens of millions of lines of code. Thousands of analysis pipelines have produced tens of thousands of potential findings, while deeper investigation has exposed dozens of flaws capable of qualifying for formal vulnerability identifiers. Twelve Common Vulnerabilities and Exposures, or CVEs, have already been assigned, while roughly another dozen flaws are moving through coordinated disclosure processes.

Among the assigned vulnerabilities are CVE-2026-13242 and CVE-2026-55803. The technology has also identified weaknesses in widely used web extensions and open-source projects, widening its potential relevance beyond investigations of proprietary enterprise software.

Rather than asking a single large language model to inspect an entire application, the system divides vulnerability discovery into a series of specialised tasks. An initial agent examines the purpose and structure of the codebase, studies available documentation and identifies directories that should be excluded from analysis, such as testing components. Other agents focus on areas including authentication, authorisation and routing.

Their findings are consolidated into a threat model that must be reviewed by a human consultant before the automated investigation proceeds. Discovery agents then analyse files for potential entry points, including HTTP routes, inter-process communication listeners and other locations where user-controlled data can enter an application.

Additional agents collect surrounding context that could determine whether apparent weaknesses are genuinely exploitable. This includes permission checks, sanitisation functions, routing conditions and nested function calls. The approach is intended to address a long-standing limitation of conventional static-analysis tools, which can identify suspicious coding patterns without always understanding whether vulnerable code can actually be reached by an attacker.

The system separately evaluates control flow and data flow. Control-flow analysis can identify weaknesses such as missing authorisation checks, privilege escalation opportunities and cross-site request forgery. Data-flow agents trace untrusted input through an application to determine whether it can reach sensitive operations capable of enabling SQL injection, cross-site scripting, command injection or path traversal.

Potential vulnerabilities then pass through several validation agents before being confirmed, rejected or marked for further examination. Verified findings are deduplicated and assigned risk levels. Security specialists subsequently attempt to reproduce the weakness and validate proof-of-concept exploits, preserving human oversight at the stage where an automated hypothesis becomes an actionable security finding.

AVDH has also demonstrated value in offensive-security simulations. During one customer engagement, analysis of a web application uncovered a remote-code-execution vulnerability that provided a route for initial access, showing how the same technology can help red teams locate weaknesses that attackers might exploit.

The development reflects a wider shift towards multi-agent security systems rather than reliance on individual AI models. Research published this year has shown that the design of the surrounding agent harness — including roles, tools, communication and verification procedures — can substantially influence vulnerability-discovery performance even when the underlying language model remains unchanged.

Google is positioning AVDH as a point-in-time tool for intensive investigations rather than continuous software monitoring. Its CodeMender technology is designed to complement that approach by scanning code, validating exploitable weaknesses and generating tested patches within development workflows.

The Agentic Vulnerability Discovery Harness, or AVDH, combines multiple specialised AI agents with structured orchestration and human security expertise. Mandiant has deployed the system for about 10 months across incident-response investigations, penetration tests, red-team exercises and proactive source-code reviews.

The result involving more than 100 critical vulnerabilities emerged during an investigation into corporate repositories stolen by attackers. The findings were classified as true positives rather than unverified alerts, an important distinction because automated security scanners can produce large volumes of findings that ultimately prove harmless. Human specialists remained involved in confirming exploitability.

AVDH has been used to examine software environments containing tens of millions of lines of code. Thousands of analysis pipelines have produced tens of thousands of potential findings, while deeper investigation has exposed dozens of flaws capable of qualifying for formal vulnerability identifiers. Twelve Common Vulnerabilities and Exposures, or CVEs, have already been assigned, while roughly another dozen flaws are moving through coordinated disclosure processes.

Among the assigned vulnerabilities are CVE-2026-13242 and CVE-2026-55803. The technology has also identified weaknesses in widely used web extensions and open-source projects, widening its potential relevance beyond investigations of proprietary enterprise software.

Rather than asking a single large language model to inspect an entire application, the system divides vulnerability discovery into a series of specialised tasks. An initial agent examines the purpose and structure of the codebase, studies available documentation and identifies directories that should be excluded from analysis, such as testing components. Other agents focus on areas including authentication, authorisation and routing.

Their findings are consolidated into a threat model that must be reviewed by a human consultant before the automated investigation proceeds. Discovery agents then analyse files for potential entry points, including HTTP routes, inter-process communication listeners and other locations where user-controlled data can enter an application.

Additional agents collect surrounding context that could determine whether apparent weaknesses are genuinely exploitable. This includes permission checks, sanitisation functions, routing conditions and nested function calls. The approach is intended to address a long-standing limitation of conventional static-analysis tools, which can identify suspicious coding patterns without always understanding whether vulnerable code can actually be reached by an attacker.

The system separately evaluates control flow and data flow. Control-flow analysis can identify weaknesses such as missing authorisation checks, privilege escalation opportunities and cross-site request forgery. Data-flow agents trace untrusted input through an application to determine whether it can reach sensitive operations capable of enabling SQL injection, cross-site scripting, command injection or path traversal.

Potential vulnerabilities then pass through several validation agents before being confirmed, rejected or marked for further examination. Verified findings are deduplicated and assigned risk levels. Security specialists subsequently attempt to reproduce the weakness and validate proof-of-concept exploits, preserving human oversight at the stage where an automated hypothesis becomes an actionable security finding.

AVDH has also demonstrated value in offensive-security simulations. During one customer engagement, analysis of a web application uncovered a remote-code-execution vulnerability that provided a route for initial access, showing how the same technology can help red teams locate weaknesses that attackers might exploit.

The development reflects a wider shift towards multi-agent security systems rather than reliance on individual AI models. Research published this year has shown that the design of the surrounding agent harness — including roles, tools, communication and verification procedures — can substantially influence vulnerability-discovery performance even when the underlying language model remains unchanged.

Google is positioning AVDH as a point-in-time tool for intensive investigations rather than continuous software monitoring. Its CodeMender technology is designed to complement that approach by scanning code, validating exploitable weaknesses and generating tested patches within development workflows.