Mandiant‘s Agentic Vulnerability Discovery Harness (AVDH) is an internal, point-in-time architecture for agentic source code review that pairs multi-agent orchestration with frontline human expertise. The authors argue that adversarial AI has accelerated the pressure on defenders: when proprietary source code leaks, attackers can deploy machine-speed AI tools, while manual review cannot keep pace. AVDH is built on the Google Agent Development Kit (ADK) and follows a strictly deterministic, waterfall-like pipeline: each phase completes before the next, producing a prioritized, risk-rated list of findings for human review. It consumes environmental inputs such as asset inventories, SBOMs, architecture documentation, and threat intelligence, which are fed into a distilled human knowledge base to let agents select relevant skills, language rules, and vulnerability patterns. The authors also note that this approach aligns with agentic orchestration capabilities now available in Google Antigravity.

The pipeline begins with an Explorer agent that identifies the codebase’s core purpose and domain, reviews documentation, flags exclusions, and dispatches Specialist Explorer subagents for authentication, authorization, routing, and similar areas. A Threat Model Synthesis agent aggregates their output, and the consultant must approve the resulting threat model before analysis continues. Next, parallelized Discovery agents use Gemini Flash Lite to extract application entry points and user-input sources. Each entry point is assigned an Enrichment agent, which gathers distributed context such as sanitizers, permissions, and routing conditions and decides whether Access Control analysis, Data Flow analysis, or both are needed.

Hypothesis generation is deliberately expansive. The Access Control agent validates security assumptions and surfaces flaws like missing authorization, privilege escalation, and CSRF. The Data Flow Analysis agent traces user input from entry points through nested calls, sanitizer transformations, and storage boundaries to dangerous sinks, exposing SQL injection, XSS, command injection, and path traversal. A consultant-configurable Confidence Filter gates the volume of hypotheses. Validation then uses multiple high-temperature Validation agents to assess each hypothesis, paired with a single Validation Synthesis agent that decides whether it is confirmed, disproven, or rejected; higher temperature broadens the reasoning considered before the final call.

Mandiant reports strong real-world results. In ten months of use, AVDH found over 100 true-positive critical vulnerabilities in two days during an incident response investigation involving stolen corporate repositories. The harness has analyzed environments with tens of millions of lines of code, executed thousands of pipelines, and produced tens of thousands of findings. That work led to 12 assigned CVEs, including CVE-2026-13242 and CVE-2026-55803, with another dozen in active disclosure. In a client engagement, the harness found a remote code execution vulnerability that enabled initial access.

Human review remains central. After the harness deduplicates and risk-rates confirmed findings, consultants dynamically replicate the exploitation and execute proof-of-concept code, discarding anything that fails. To address false negatives, Mandiant engineered a rules-based layer that injects consultant expertise via hierarchical prompts: software domain at the top, followed by language, framework, and vulnerability rules. Language and framework rules apply across the pipeline; vulnerability rules apply during discovery, validation, and risk-rating.

For benchmarking, Mandiant built proprietary synthetic codebases because public datasets carry training-data contamination risks. Every injected vulnerability is manually verified to be reachable and dynamically exploitable. A dedicated Grading agent compares pipeline findings to ground truth, requiring precise vulnerability matches rather than loose semantic similarity. Secondary agents handle false-positive triage and duplicate clusters, then a human expert reviews the AI judges. This cycle runs across multiple domains and architectures for every major release, with results averaged to account for LLM non-determinism.

The authors frame AVDH as a defense multiplier: automating routine vulnerability discovery frees consultants to focus on complex flaws. They position AVDH for point-in-time assessments and deep proactive discovery, and describe CodeMender as a complementary continuous scanning layer; together, they form a two-layered defense, with Google AI Threat Defense offering an always-on platform that includes CodeMender. The article also advertises a Cyber Defense Summit presentation in September 2026.