Agentic Source Code Review for Vulnerability Discovery

What it is

Mandiant's public write-up of AVDH's architecture — shared for the first time so defenders can build similar pipelines in-house. Beyond the tool description, the post gives design specifics and operational scale that the summary omits: AVDH has been in use for 10 months, has analyzed environments spanning tens of millions of lines of code, and run thousands of pipelines producing tens of thousands of findings.

How it works

The value the page argues for an LLM over traditional pattern-matching engines: it can distinguish code reachable by a standard user from admin-restricted code, and from code that never executes — i.e. reachability and business-logic context, not just rule hits.

The pipeline is threat-model-first. An Explorer agent identifies the codebase's core purpose and software domain, reviews documentation, flags directories to exclude (e.g. unit tests), and dispatches Specialist Explorer subagents scoped to authentication, authorization, routing, and other categories. A Threat Model Synthesis agent aggregates their output. A human approval gate — with textual and visual representations of the threat model — must pass before the rest of the pipeline runs, so downstream analysis rests on a verified foundation.

Preconditions / Affected

For this approach to work, the harness needs rich environmental inputs fed into a distilled human knowledge base: asset inventories, SBOMs, architecture documentation, and threat intelligence. These let agents dynamically select relevant skills, language rules, and vulnerability patterns. Without them the threat model loses attack-vector, business-logic, and reachability context. Built on Google's Agent Development Kit (ADK); the post ties it to Google Antigravity as the orchestration workspace.

What to do

This is a design pattern to replicate, not a released tool — there are no commands or patches. If building something similar:

• Establish and human-validate a threat model before running vulnerability analysis; treat the approval gate as mandatory.
• Front-load organizational context (SBOMs, asset inventory, architecture docs, threat intel) rather than pointing agents at raw source alone.
• Pair with continuous scanning: Mandiant positions AVDH alongside CodeMender as a two-layered defense (point-in-time deep review + ongoing scanning).

Caveats

This is a vendor post about an internal, "point-in-time" architecture — not a product you can download, and results (100+ true-positives in two days, RCE enabling initial access on a client web app, the CVE counts) are self-reported by Mandiant. The specific agent skills, language rules, and prompts are not disclosed; the post gives the orchestration shape and required inputs, not a reproducible implementation.