Staying Ahead of Adversarial AI Through Agentic Source Code Review

Staying Ahead of Adversarial AI Through Agentic Source Code Review
Mandiant describes AVDH, an agentic vulnerability discovery harness that combines LLMs, structured orchestration, and human expertise to find and validate software flaws at scale. In real-world use, it uncovered over 100 critical vulnerabilities in two days, contributed to 12 assigned CVEs, and identified an RCE in a client web application source code. #AVDH #Mandiant #CodeMender #GoogleAntigravity #GoogleAgentDevelopmentKit #GeminiFlashLite #CVE-2026-13242 #CVE-2026-55803

Keypoints

  • AVDH is a structured, multi-agent vulnerability discovery framework designed for proactive code analysis, penetration testing, red teaming, and incident response.
  • The system combines AI-driven analysis with human subject-matter expertise to reduce false positives and improve validation quality.
  • In a recent incident response case involving stolen corporate repositories, AVDH found more than 100 true-positive critical vulnerabilities in two days.
  • The harness has been used at scale across tens of millions of lines of code and thousands of pipelines, generating tens of thousands of findings.
  • Its use has led to 12 assigned CVEs, including CVE-2026-13242 and CVE-2026-55803, with additional disclosures still in progress.
  • The pipeline includes threat modeling, entry point discovery, context enrichment, hypothesis generation, validation, and expert review to systematically assess vulnerabilities.
  • Google AI Threat Defense and CodeMender are presented as complementary continuous-defense tools that can be paired with AVDH for a two-layered strategy.

MITRE Techniques

  • [T1190 ] Exploit Public-Facing Application – The harness found a remote code execution flaw in a client web application that enabled initial access (‘quickly found a remote code execution (RCE) vulnerability that enabled initial access’).
  • [T1059 ] Command and Scripting Interpreter – The analysis specifically looks for command execution paths where malicious input could reach a dangerous sink and execute (‘command injection’).
  • [T1083 ] File and Directory Discovery – The Explorer agent reviews directories and excludes areas such as unit tests while mapping the codebase (‘flags directories to exclude from scanning’).
  • [T1068 ] Exploitation for Privilege Escalation – The Access Control agent identifies flaws such as missing authorization and privilege escalation (‘including missing authorization, privilege escalation’).
  • [T1003 ] OS Credential Dumping – Not mentioned directly in the article; no technique mapping applicable.
  • [T1056.001 ] Input Capture: Keylogging – Not mentioned directly in the article; no technique mapping applicable.
  • [T1195 ] Supply Chain Compromise – The article focuses on stolen corporate repositories and open-source code analysis, but does not explicitly describe supply chain compromise execution; no direct technique mapping is claimed.
  • [T1021 ] Remote Services – The framework analyzes IPC listeners and HTTP routes as attack vectors (‘HTTP routes, inter-process communication (IPC) listeners’).
  • [T1595 ] Active Scanning – The harness performs broad, parallelized code scanning across in-scope files to discover entry points and vulnerabilities (‘parallelized Discovery agents to analyze every in-scope file’).
  • [T1040 ] Network Sniffing – Not mentioned directly in the article; no technique mapping applicable.
  • [T1211 ] Exploitation for Defense Evasion – Not mentioned directly in the article; no technique mapping applicable.
  • [T1190 ] Exploit Public-Facing Application – The framework identifies CSRF, XSS, SQL injection, path traversal, and related flaws in web applications (‘cross-site request forgery (CSRF)’, ‘SQL injection, cross-site scripting (XSS), command injection, and path traversal’).

Indicators of Compromise

  • [CVE IDs ] Assigned and in-disclosure vulnerabilities identified through the harness – CVE-2026-13242, CVE-2026-55803, and 12 additional CVEs in active disclosure.
  • [Software/tools ] Security and AI platforms referenced in the workflow – AVDH, CodeMender, Google Antigravity, Google Agent Development Kit (ADK), and Gemini Flash Lite.
  • [Organizations ] Entities and environments involved in the analysis – Mandiant, Google AI Threat Defense, and client/stolen corporate repositories.
  • [Dates / events ] Public presentation and deployment timeline references – Cyber Defense Summit September 15-16, 2026, and a recent incident response investigation.


Read more: https://cloud.google.com/blog/topics/threat-intelligence/staying-ahead-of-adversarial-ai-through-agentic-source-code-review/