Google DeepMind launches CodeMender agent for AI code security

Google DeepMind has unveiled CodeMender, an autonomous agent that has been designed to find, fix and help prevent software vulnerabilities at scale.

Announced on 6 October 2025, the system uses Google’s Gemini models to reason about code and propose validated patches, with all changes receiving human review before submission upstream.

DeepMind has said CodeMender combines root‑cause analysis with “self‑validated” patching. It has deployed multiple tools static and dynamic analysis, fuzzing, differential testing and SMT solvers alongside a multi‑agent critique stage that evaluates whether a proposed change is correct, secure and regression‑free before surfacing it for human sign‑off.

The approach has been built on the reasoning capabilities of recent Gemini models.

What has CodeMender achieved so far?

In its first six months, CodeMender has upstreamed 72 security fixes to open‑source projects, including codebases of up to 4.5 million lines. The team has emphasised that every patch has been validated automatically across multiple dimensions and then reviewed by researchers prior to submission.

Beyond patching discovered flaws, the agent has been applied to proactively rewrite code to make entire classes of bugs harder to exploit.

In one example, CodeMender has added -fbounds-safety annotations to parts of the widely used libwebp library targeted by CVE‑2023‑4863 so that compiled code includes bounds checks designed to neutralise buffer overflows.

The launch has arrived alongside Google’s broader AI security push. The company has introduced an AI Vulnerability Reward Program and updated its Secure AI Framework to SAIF 2.0, focusing on risks and controls for agentic systems. Together with CodeMender, Google has positioned these as measures that have aimed to speed time‑to‑patch and strengthen open‑source resilience.