Google CodeMender on Gemini Enterprise Agent Platform is an autonomous, AI-powered security agent originally developed by Google DeepMind to find, verify, and remediate deep vulnerabilities in your codebase. Tested across Google's own production environments, it uses a fine-tuned security harness that is built and continuously updated by Google DeepMind, and a selection of models to analyze root causes, run exploit simulations to confirm vulnerabilities are real, and generate precise code fixes.

Google AI Threat Defense is an always-on, autonomous security platform designed to outpace AI-driven attacks. It seamlessly integrates the reasoning power of frontier models like Gemini, the contextual cloud-risk prioritization of Wiz, the autonomous code remediation capabilities of CodeMender, and the frontline incident response expertise of Mandiant. Within AI Threat Defense, CodeMender serves as the primary code security agent, bridging real-world cloud exposures directly with automated, developer-approved codebase patching.

CodeMender shifts security from passive scanning to active remediation. Unlike traditional static scanners that flood teams with false positives, or standalone models lacking execution logic and context, CodeMender pairs advanced AI models with a fine-tuned harness and comprehensive security toolbox to deliver better performance in vulnerability fixing than models alone. CodeMender enables you to:

• Deploy the best-fit model. You can choose from multiple models to optimize for costs, speed, deep scanning, and coding performance.
• Accelerate machine-scale remediation. You can now eliminate remediation bottlenecks caused by manual verification and patching, while keeping developers in the loop.
• Prioritize fixes by exploitability. You can run proof-of-concept exploits and execute simulations to verify vulnerabilities in the code are exploitable, and prioritize tokens on fixing critical ones first.

Traditional static scanners rely on simple pattern matching, which generates high volumes of noisy alerts and false positives. CodeMender acts differently, performing high-signal, deep scans across complex, multi-language codebases. It leverages advanced program analysis to trace control flows and data paths, allowing it to understand the unique context, goals, and functionality of your application to discover sophisticated vulnerabilities that static and model-only scanning miss.

The most critical bottleneck in application security is the time spent triaging false positives. CodeMender solves this through execution-based verification. Rather than statically analyzing code patterns, the agent builds your code and executes a secure proof-of-concept exploit. If the exploit succeeds, the vulnerability is verified to be exploitable, allowing your security practitioners and developers to focus their energy and token costs on legitimate risks.

Once a security flaw is verified, CodeMender automatically generates a patch, delivered as a standard code diff directly within developer tools. To ensure the fix is production-ready, CodeMender uses an LLM-as-a-judge framework to validate the patch. The agent automatically tests the modified code to ensure program equivalence, meaning the patch resolves the root cause of the vulnerability without altering the application's intended behavior or breaking existing features. The patch is presented to the developer as a code diff for them to review and approve before committing the patch to their code base.

For the latest list of supported models, you can check our documentation.  CodeMender with Gemini 3.5 Flash Cyber will be exclusively available to a small set of governments and trusted partners. We plan to expand this access over time.

CodeMender is built for enterprise-scale complexity, providing comprehensive coverage across major language ecosystems shown below.  For the latest, you can check our documentation.

• Languages: C/C++, Java, Python, TypeScript/JavaScript, Go, Rust, and Ruby
• Frameworks: Broad support for standard libraries within these languages, as well as common enterprise frameworks (e.g., HTML/CSS, Django, Flask, React, Spring Boot, and Express)

CodeMender is hosted as a secure headless agent on the Gemini Enterprise Agent Platform. Developers interact with it using CodeMender’s local Command Line Interface (CLI) and tools such as VS Code to run scans, review findings, trigger exploit verification, generate patches, and manage sessions. Core reasoning and orchestration occur in the cloud, while code compilation, test execution, and exploit simulations run in customer-managed local sandboxes or isolated VMs.

Yes. In addition to interactive developer workflows, CodeMender can operate as a headless agent. For automated pipelines, the CodeMender binary can be packaged inside a custom container.

Under the Google AI Threat Defense system CodeMender works with Wiz:

• Context-Rich Scanning and Prioritization:  Wiz orchestrates agentic application security, analyzing applications to prioritize investigations. It calls CodeMender to scan code (coming soon), enriches findings within the Wiz Security Graph with deployment context, and triggers Wiz Red Agent for AI pentesting to prove exploitability, ensuring teams focus on the highest-risk vulnerabilities.
• Targeted Remediation:  Wiz serves as a command center for governing and scaling remediation. The Wiz Green Agent orchestrates this lifecycle by directing CodeMender to generate and test high-fidelity patches enriched with application context from the Security Graph. This workflow empowers teams to resolve complex vulnerabilities with unprecedented speed and precision.

Absolutely. CodeMender enforces a strict human-in-the-loop governance model. The agent does not automatically apply patches to your version control system or push code directly to production repositories. All suggested patches are provided as local code diffs, and changes are only committed to the repository after a developer has manually reviewed and approved them.

Your proprietary source code remains entirely private and secure. CodeMender is architected from the ground up with strict data privacy perimeters:

• No Repository Uploads:  You never need to upload your source repositories to Google Cloud. The CodeMender agent only requests and processes the minimum necessary code snippets via secure communication with the CLI client on your local machine.
• Sandbox Compilation and Verification:  To ensure logic remains on the client side, all code builds, test executions, and exploit simulations are run entirely on customer-managed infrastructure or sandboxes — never on Google-managed infrastructure.
• Minimal Data Retention: Google does not retain any source code, prompts, scan outputs, or patches on Google-side storage beyond active session data, which is stored encrypted for up to seven days from session creation to enable resumption of long-running find and fix scans. Scan outputs, exploits, and patch code are sent back to the client and stored locally on your machine.
• No Model Training: Your proprietary code, prompts, and security findings are never used to train Google's foundation models.

Hosted securely on the Gemini Enterprise Agent Platform, CodeMender provides the following enterprise guardrails:

• VPC Service Controls (VPC-SC):  Establishes strict network perimeters around agent communication, providing enhanced protection against data exfiltration.
• Secure Traffic Routing:  Gives customers the option to route all CLI and agent communications exclusively through their VPC on Google's private, encrypted backbone network, mitigating eavesdropping risks.