Google has introduced CodeMender, an artificial intelligence agent designed to identify security vulnerabilities in code, confirm their exploitability, and generate patches for developer review. The company states that this tool is a direct response to the increasing use of AI by attackers to accelerate their operations, emphasizing the need for automated defensive measures operating at a comparable speed.
CodeMender is currently available in preview through the Gemini Enterprise Agent Platform, leveraging Google's generally available Gemini models. It is also offered as a component of AI Threat Defense, Google's AI-powered security platform. A specialized version, integrated with Gemini 3.5 Flash Cyber, is initially restricted to a select group of governments and trusted partners, with plans for broader access in the future. Google also intends to support third-party frontier models later in the year.
Google highlights that CodeMender's underlying framework is continuously updated with the latest research from Google DeepMind, incorporating current agent skills, security tools, and system prompts. The agent operates within the secure-by-design Agent Platform, benefiting from enterprise-grade governance and security features such as secure traffic routing via a Virtual Private Cloud (VPC), data isolation and encryption, and a policy of zero retention for source code data.
The agent's operation is structured in three distinct stages. First, CodeMender scans code repositories for vulnerabilities that static analysis tools frequently overlook. These include memory corruption, various injection flaws, web security issues, cryptographic weaknesses, and insecure data handling practices. The tool supports a range of programming languages, including C/C++, Go, Java, Python, Ruby, Rust, and TypeScript.
In its second stage, CodeMender aims to validate the exploitability of identified flaws. Rather than merely matching patterns, the agent constructs a proof-of-concept exploit and executes it within a customer-controlled sandbox environment. Google asserts that this step significantly reduces the incidence of false positives.
Upon confirmation of a vulnerability, CodeMender proceeds to write a patch. This proposed fix is then evaluated by a separate model acting as a judge, which is designed to detect instances where the patch might inadvertently introduce new issues or break existing application functionality. The generated patch is presented to developers as a code diff, and no changes are committed to a repository without explicit manual approval. When deployed within AI Threat Defense, CodeMender integrates with Wiz to contextualize vulnerabilities within deployment environments and initiate automated penetration testing.
