AVDH acts as a force multiplier, automating routine vulnerability discovery so defenders can focus on complex exploit chains, business-logic flaws, and adversarial activity that require human judgment.

Google’s Mandiant has unveiled the Agentic Vulnerability Discovery Harness (AVDH), an AI-powered framework that uses a chain of specialized agents guided by human review to hunt for exploitable flaws in large codebases, including a live investigation where it found over 100 verified, high-severity vulnerabilities in two days.

Final validation rests with human experts who reproduce exploit paths and test proof-of-concept code to confirm risk, discarding findings that fail validation.

The framework incorporates distilled human knowledge through structured rule sets across languages, frameworks, and vulnerability types to guide analysis and reduce misses.

Discovery and enrichment map targets, identify entry points, isolate user-controlled input, and assess whether exposed functions have weak checks or if attacker data can reach dangerous execution sinks.

AVDH outputs a risk-rated set of potential flaws for consultants to investigate, reproduce, and disclose, with Mandiant applying it across environments containing tens of millions of lines of code over the past year.

Leadership notes emphasize embedding frontier models within an expert harness to automate routine vulnerability discovery.

The tool operates as a pipeline of agents built on Google’s Agent Development Kit, moving through threat modeling, entry-point discovery, context enrichment, hypothesis generation, and hypothesis validation, with human experts confirming flaws via PoC code.

Human-in-the-loop validation remains essential, with defenders advised to manually validate AI findings and maintain collaboration between humans and AI tooling.

The harness operates as a sequential pipeline rather than a single prompt, starting with threat modelling, moving through file-wide entry-point discovery, and then deeper control-flow and data-flow analyses.

Hypotheses are generated by specialized agents focused on access control and data flow, with independent validation and a synthesis agent determining evidence strength for each potential flaw.

To validate performance without training-data contamination, Mandiant created synthetic vulnerable codebases instead of relying on public vulnerability datasets.