Google’s AI review finds 100 critical flaws in two days

What this is about

Google Threat Intelligence described the architecture of its internal Agentic Vulnerability Discovery Harness, or AVDH, for the first time on August 18, 2026. Mandiant uses the system in code reviews, penetration tests, and incident response. Its most striking documented engagement followed the theft of corporate repositories: Google says AVDH found more than 100 true-positive critical vulnerabilities in two days.

That number matters because stolen source code gives attackers a detailed map. Automated tools can search it for exploitable mistakes at high speed. Defenders need to complete the same work faster without wasting time on invented findings.

What AVDH actually does

AVDH is not one chatbot and it is not a button that automatically makes software safe. It divides source-code review into a fixed sequence of specialized agents. An initial agent creates a threat model: What kind of application is this, which parts are reachable, and which directories should be excluded? A human approves that model.

The pipeline then inventories entry points such as HTTP routes and interfaces. Other agents collect surrounding code, permission checks, and data flows. Separate stages look for missing authorization, SQL injection, cross-site scripting, command injection, and path traversal. Several validation agents test the hypotheses before people reproduce an exploit. Findings that cannot be confirmed are discarded.

Google built the orchestration with its Agent Development Kit and uses rules maintained by Mandiant specialists for programming languages, frameworks, and vulnerability classes. Over ten months, the vendor says AVDH analyzed environments containing tens of millions of lines of code. The reviews produced twelve assigned CVEs, while roughly another dozen were in coordinated disclosure on August 18.

Why it matters

The important part is not the word “agent” but the combination of automation, fixed process logic, and human verification. Conventional scanners are good at known patterns but can struggle with business logic and complex authorization paths. A language model can connect more context, yet it can also make false assumptions without controls. AVDH tries to contain both weaknesses through sequential checks.

For development teams, this shifts the bottleneck. If machines produce thousands of suspected findings, reliable confirmation becomes more valuable than raw volume. The CVEs cited by Google, including CVE-2026-13242 and CVE-2026-55803, make part of the output publicly checkable. The system’s overall performance has not been independently benchmarked, however.

In plain language

A manual code review is like looking for leaks in a large building. One helper runs from room to room and reports every damp patch. AVDH works more like a team: one person reads the plans, others inspect pipes and access points, several check the suspected leaks, and only then does a plumber confirm that water is actually escaping.

A practical example

Suppose a retailer loses a repository containing 500,000 lines of code. Its security team must learn whether public routes can reach private customer data before an attacker does. AVDH might first inventory 2,000 entry points, narrow those to 80 plausible data flows, and send 15 hypotheses to validation agents. If dynamic testing confirms three reproducible flaws, people prioritize those three by reachability and potential harm.

The numbers in this example are fictional. The workflow illustrates the practical value: automation reduces the search space while people verify the claim that matters.

Scope and limits

• The figure of more than 100 critical flaws comes from Google and concerns an unnamed customer engagement. The codebase, comparison effort, and error rate are not public.
• AVDH is an internal Mandiant system, not a generally available product. Other teams can copy the published architecture, but they do not automatically receive the same rules, models, or expertise.
• The process needs complete, trusted source code and human exploit validation. It does not replace runtime defenses, secure development, least privilege, or fast patching.

SEO & GEO keywords

Google Threat Intelligence, Mandiant, AVDH, Agentic Vulnerability Discovery Harness, AI code review, source code security, vulnerability discovery, CVE, DevSecOps, human validation