Agentic Code Review: How Mandiant Stays Ahead of Adversarial AI

When adversarial AI starts sniffing around your source code, you’d better have more than a rubber duck and a prayer. Mandiant’s new Agentic Vulnerability Discovery Harness (AVDH) is like a team of capybaras with magnifying glasses—calm, methodical, and surprisingly deadly at finding bugs before the caimans do.

In the swamp of modern cybersecurity, attackers are using machine-speed AI to exploit stolen code. Mandiant’s answer? Fight fire with fire, but with a structured, human-guided approach that keeps the AI on a leash. They’ve shared the internals of AVDH for the first time, and it’s a game-changer for defenders.

Real-World Results: Numbers That Bite

AVDH isn’t just a lab experiment. In the last 10 months, it’s been tearing through codebases with impressive results:

• 100+ critical vulnerabilities found in just two days during an incident response.
• Analyzed environments with tens of millions of lines of code.
• Generated tens of thousands of findings across thousands of pipelines.
• Led to 12 assigned CVEs, including CVE-2026-13242 and CVE-2026-55803.

That’s the kind of speed that makes manual review look like a sloth on a coffee break. But it’s not just about speed—it’s about precision and depth.

Architecting the Pipeline: A Waterfall in a Swamp

AVDH is built on the Google Agent Development Kit (ADK), which orchestrates specialized agents in a sequential pipeline. Think of it like a waterfall: each phase must complete before the next begins. This isn’t a chaotic free-for-all; it’s a structured flow that mirrors how Mandiant’s consultants work in the field.

The pipeline starts with Threat Modeling, where an Explorer agent maps out the codebase’s purpose, identifies key areas, and dispatches Specialist Explorers to dive into authentication, authorization, routing, and more. The result is a visual threat model that human consultants approve before moving on.

Next, Entry Point Discovery uses parallelized agents to scan every file for HTTP routes, IPC listeners, and other attack surfaces. Then, Context Enrichment agents gather surrounding code to understand how each entry point fits into the bigger picture—because vulnerabilities often hide multiple hops away.

Hypothesis Generation and Validation: The Brainstorming Phase

Here’s where it gets interesting. The Access Control and Data Flow Analysis agents generate hypotheses about potential vulnerabilities. They’re encouraged to think broadly, with a Confidence Filter keeping the noise down.

Then, multiple Validation agents with high temperature settings—think of it as creative chaos—assess each hypothesis. A Synthesis agent then weighs their verdicts to confirm, disprove, or reject the finding. This multi-perspective approach reduces false positives and catches more real issues.

Human Expertise: The Secret Sauce

AI is powerful, but it’s not infallible. That’s why Mandiant injects human expertise at two critical points:

• Expert Validation: After the harness finds something, human consultants manually verify by exploiting the vulnerability and running proof-of-concept code. If it doesn’t hold up, it’s discarded.
• Distilled Knowledge: They’ve built a rules-based system that encodes Mandiant’s collective wisdom about languages, frameworks, and vulnerability patterns. This knowledge is injected into the pipeline, guiding agents to look for the right things.

This human-in-the-loop approach ensures that the AI isn’t just hallucinating bugs—it’s finding real, exploitable flaws that matter.

Measuring Success: Benchmarking Against the Swamp

You can’t improve what you can’t measure. Mandiant built a proprietary benchmark suite with synthetic codebases that mimic real-world applications. Every vulnerability is manually verified to be reachable and exploitable. They grade the pipeline’s output with a dedicated Grading agent, plus human review, to ensure accuracy.

This rigorous testing cycle runs for every major release, averaging results to account for LLM non-determinism. It’s a solid approach for anyone building similar tools.

Conclusion: Defenders Can Win

Manual code review can’t keep pace with AI-driven attacks. But with a harness like AVDH, defenders can automate the discovery of routine vulnerabilities, freeing humans to focus on complex flaws. It’s a practical multiplier for human expertise.

For continuous monitoring, Mandiant suggests pairing AVDH with CodeMender for a two-layered defense. And if you want to see it in action, catch their talk at the Cyber Defense Summit in September 2026.

In the ever-shifting swamp of cybersecurity, it’s good to know the capybaras are armed with AI-powered bug-finding skills. Stay ahead of the caimans, folks.

Original announcement published on Google Cloud.