Mandiant Opens Agentic Security Harness to Industry

Artificial Intelligence & Machine Learning
,
Next-Generation Technologies & Secure Development
,
The Future of AI & Cybersecurity

AVDH Scans Enterprise Code at Scale to Find and Validate Exploit Paths

Emilia David •
August 21, 2026    

See Also: OnDemand | Security Operations in the Age of AI

The Agentic Vulnerability Discovery Harness analyzes code to quickly identify exploit paths during reviews, penetration testing and red team operations. As a harness, it adds a deterministic set of skills to the agent.

Mandiant said in a blog post that since using AVDH, it “greatly accelerated how Mandiant discovers vulnerabilities at scale.” The company said the harness helped “dozens of assignable flaws,” and it used it to analyze “tens of millions of lines of code and execute thousands of pipelines.”

Alex Tselevich, senior consultant at Google Mandiant, told ISMG in an interview that the speed at which agentic risk is rising necessitates newer ways to find and plug exploits.

“The entire intention of this research release is to share what we found that works and share some of that research so that the rest of the industry and other security teams can benefit from it,” Tselevich said.

The Mandiant team said the speed at which the attackers are using AI tools means defenders also need better tooling to protect themselves. Tselevich said they wanted to try using harnesses because more traditional source code review tooling has limitations in accuracy and noise. He said a single-agent approach also could not work for their needs because of model context sizes, the lack of programmatic validation and the need to stick to a complex workflow.

AVDH is vendor agnostic, though the Mandiant team said the harness can run alongside Google’s CodeMender scanning tool.

The harness approach, Tselevich said, made the most sense because it let the team combine the flexibility of agents and AI models with programmatic checks that can discard false positives without overwhelming the agents.

Mandiant technical manager Michael Maturi told ISMG that if other labs want to take on a similar harness, they need first to understand what they want to protect. For Mandiant, they focused on the most critical code first and pointed the agent and harness toward pieces of the code base that provide access to sensitive systems and data and have API endpoints.

The team used Google’s Agent Development Kit to build out the harness. AVDH runs in several stages: Threat Modeling, Entry Point Discovery, Context Gathering, Hypotheses Generation and Hypothesis Validation. Mandiant takes an additional step: human subject-matter experts validate what the harness discovers, and a human prepares the information for formal disclosure.

The threat modeling pipeline starts by deploying an Explorer agent, which will identify the codebase’s purpose. It then spins up other specialist Explorer agents that focus on areas like authentication, authorization, routing and other domain-specific categories before passing its findings to a Threat Model Synthesis agent.

Once the threat model is established, AVDH moves on to finding an entry point, which employs Discovery agents that isolate and extract all sources of user input. The harness then assigns the entry points to dedicated Enrichment agents that look deeper into other critical components each code needs, like permissions or routing conditions. It determines whether the entry points need more information about access control or data flow.

AVDH then starts generating a hypothesis that evaluates protections around each entry point and validates security processes. It also checks whether access is restricted or whether it was inadvertently exposed to unauthorized users. This step also tracks the flow of user inputs and follows where data travels.

The last step is Hypothesis Validation, where the agent evaluates the reasoning behind how secure the entry point is. At the end of the five stages, AVDH would have determined if a potential vulnerability is confirmed, disproven or rejected. These results are passed on to the human experts for validation.

Tselevich said that existing red teams around security, but he said it is difficult to scale many of these needs. He said the experiments, and the codebase Mandiant pointed AVDH to, grew the harness’s maturity to enterprise scale.

“We are giving you a blueprint that we, at scale, have been using to scan tens of millions of lines of code from enterprise projects of basically every scale,” he said.