“It will support third-party frontier model options later this year,” Google said in a statement on July 21.
DeepMind's CodeMender: from research project to managed enterprise agent
CodeMender, developed by Google DeepMind and first introduced in October 2025 as an advanced security research project, is now available as a fully managed AI code security agent inside Google Cloud. In its research phase, the system combined Gemini reasoning models with static and dynamic analysis to find, diagnose and patch software defects; it also produced 72 upstreamed security fixes to major open-source projects under human-in-the-loop oversight. Google has moved the tool from laboratory work to an enterprise-grade service designed to operate inside customer development pipelines and cloud environments.
How CodeMender verifies and delivers fixes
The managed agent does more than flag theoretical issues. CodeMender will build and run proof-of-concept exploits in customer-managed sandboxes to test whether vulnerabilities are truly exploitable, then generate tested fixes and deliver them as code differences for developer review. Google also adds a large language model–as–a–judge mechanism that evaluates proposed changes against organizational rules and checks that patches do not break core business logic. The agent is multi-model, letting developers select among Gemini 3.5 Flash (the default), Gemini 3.1 Pro and Gemini 3 Flash to balance cost, speed and depth of scanning.
Integration with Google Cloud, Gemini Enterprise Agent Platform, and AI Threat Defense
CodeMender is available in public preview on the Gemini Enterprise Agent Platform and can be accessed through Google Cloud. It integrates into local developer environments via a command line interface and tools such as VS Code and the Antigravity desktop app. Organizations can also deploy CodeMender as a core component of Google AI Threat Defense, where it operates alongside Mandiant’s frontline expertise and Wiz’s contextual risk prioritization. Google said additional interoperability is planned, including the capability for Wiz to call CodeMender to scan code so Wiz’s Green Agent can orchestrate remediation.
Security controls, language support and limits on powerful tooling
Google is shipping several operational controls for enterprise customers: secure traffic routing through a customer’s Virtual Private Cloud, data isolation, encryption and zero retention of source code data. At preview, CodeMender supports scanning and remediation across C, C++, Go, Java, Python, Ruby, Rust and TypeScript/JavaScript, and recognizes common enterprise frameworks including Django, Flask, React, Spring Boot and Express. Google also disclosed a highly specialized security-tuned model — Gemini 3.5 Flash Cyber — that is being piloted with limited access and restricted “exclusively to governments and trusted partners” because of its powerful dual-use capabilities.
What this means for technologists, policymakers, and open-source maintainers
- Technologists and security teams: Teams will be able to integrate CodeMender directly into development toolchains and run exploit proof-of-concepts in isolated sandboxes; they will also need to choose among Gemini model options to match scanning depth and cost.
- Policymakers and regulators: The limited-access pilot for Gemini 3.5 Flash Cyber and Google’s planned rollout of governance features — user registration and identity management, advanced observability and audit logs, and localized data residency controls — will be points for oversight and procurement policy discussions.
- Open-source maintainers: The research phase already resulted in 72 upstreamed security fixes under human-in-the-loop review; maintainers will see a managed agent that can propose and test patches but that depends on human review before changes are upstreamed.
Google’s move places a research-proven autonomous patching capability inside a managed enterprise service with sandboxed exploit testing, an LLM judging layer and selectable Gemini models. The company has signaled continued expansion — including third-party frontier model options later this year and planned governance controls — while keeping some powerful capabilities tightly restricted. Whether those future governance features and third-party model integrations arrive on schedule, and how organizations will balance automation against oversight, are the concrete next steps set by today’s rollout.
Source: https://www.infosecurity-magazine.com/news/google-codemender-available-ai/
