Critical CVSS 10.0 Vulnerability in Google's Agent Development Kit

By Carlos Montiel | Enterprise AI Specialist
Leer en español →
Published: 2026-09-11 | By: Carlos Montiel | Reading time: ~4 minutes

On September 9, 2026, CVE-2026-79696 was disclosed: a critical vulnerability with a CVSS score of 10.0 in Google's Agent Development Kit (ADK) for Python — the open-source framework with over 90 million downloads that many companies now rely on to build and deploy AI agents.

What the vulnerability actually allows

The flaw affects adk-python versions 2.0.0 through 2.6.0, in environments where pytest is installed (OSS, Cloud Run, and GKE). It stems from improper handling of disallowed input (CWE-184) in the "adk web" component: an unauthenticated remote attacker can achieve arbitrary code execution through a manipulated test session replay. In practical terms, this is about as severe as a software vulnerability gets — which is why it carries the maximum CVSS score of 10.0.

CVE-2026-79696 technical summary: - Component: Google Cloud Agent Development Kit (ADK) for Python - Affected versions: 2.0.0 - 2.6.0 - Precondition: pytest installed in the environment (OSS, Cloud Run, GKE) - Type: improper input validation (CWE-184) in "adk web" - Impact: unauthenticated remote code execution - CVSS: 10.0 (critical) - Disclosed: September 9, 2026

Not an isolated incident for ADK

This disclosure lands just months after Pillar Security researchers documented, in June 2026, what they described as the first real-world case of one AI agent being used to exploit another — with the vulnerability residing in that same ADK for Python open-source repository. Google fixed that earlier issue in the repository, though it classified it as ineligible for a bounty since it involved social engineering. The pattern emerging here is that an agent framework as widely adopted as ADK is becoming a recurring target for both security research and real-world exploitation.

Recommended action if your team uses ADK: immediately check which version of adk-python is running in production and whether pytest is installed in those same environments (Cloud Run, GKE, or on-prem). Google has already published the official advisory for this CVE; update to the latest patched release in the adk-python repository before exposing any "adk web" instance to untrusted traffic, and avoid having pytest installed in production environments as a general practice.
What this means for a company buying or building on AI: agent frameworks — ADK, LangGraph, and their peers — are software like any other, with real attack surface and real CVEs, not some "intelligent" layer exempt from basic software security hygiene. If your company builds agents on ADK or a similar framework, treat it like any other critical dependency: keep a version inventory, patch quickly against critical CVEs, and audit which development-only components (like pytest) end up installed in production by accident. In several cases, the pace of adoption of these frameworks is outrunning the maturity of their security processes.
Carlos Montiel
Enterprise AI Solutions Architect
Specialist in LLMs, Agents and Orchestration
guatemalia.com/#contacto · info@guatemalia.com

Need to implement AI at your company?

Carlos Montiel is an enterprise AI solutions architect. He implements LLMs, agents, RAG, and orchestrators for companies across Guatemala and Latin America. Get in touch for a consultation.

Contact Carlos Montiel

info@guatemalia.com