Google has unveiled CodeMender, a managed AI security agent designed to identify, validate, and remediate software vulnerabilities at machine speed.
Announced in preview on July 22, 2023, the tool is available through the Gemini Enterprise Agent Platform and can also function as a core component of Google’s AI Threat Defense offering.
This launch comes as organizations face increasing pressure to protect their software supply chains against AI-accelerated exploitation and to reduce the time between discovering a flaw and deploying a secure fix.
CodeMender AI Agent for Automated Vulnerability Detection
Unlike conventional static analysis tools that often generate high volumes of unverified findings, CodeMender is built around a workflow that scans, verifies, and remediates vulnerabilities.
The agent scans repositories for vulnerable code, produces proof-of-concept exploits in an isolated, customer-managed sandbox to establish exploitability, and then generates tested patches for developer review.
Google states that this process allows teams to prioritize validated vulnerabilities, rather than expending resources on theoretical weaknesses or false positives.
CodeMender’s scanning capabilities are continuously updated with research from DeepMind, agent skills, security tools, and system prompts.
It is designed to identify challenging vulnerability categories, including memory corruption, code injection, web application weaknesses, cryptographic implementation flaws, and insecure data handling.
Feature Overview
| Capability | Technical Details | Security Impact |
|---|---|---|
| AI-powered scanning | Analyzes repository context, application logic, and common vulnerability classes | Detects issues traditional static and model-only tools may miss |
| Exploit verification | Builds and runs proof-of-concept exploits in a customer-managed sandbox | Reduces false positives and validates real-world risk |
| Automated remediation | Generates patches as code diffs and tests them before submission | Accelerates mitigation while preserving developer approval |
| Multi-model support | Lets organizations select models based on cost, speed, deep scanning, and coding performance | Enables flexible deployment and operational optimization |
| Language coverage | Supports C/C++, Go, Java, Python, Ruby, Rust, and TypeScript | Covers common enterprise and open-source development stacks |
| Enterprise controls | Uses VPC traffic routing, source-code isolation, encryption, and zero data retention | Helps protect sensitive proprietary code during analysis |
| Workflow integration | Connects to CI/CD pipelines, local environments, a CLI client, VS Code, and Antigravity | Fits into existing software development workflows |
By incorporating repository-level context, the agent aims to understand how an application functions instead of relying solely on vulnerable code patterns.
The verification stage is central to the product’s security proposition. CodeMender can create an exploit plan, develop exploit code, and execute it within an isolated sandbox controlled by the customer.
Security teams can then use the results to determine whether a finding is genuinely exploitable and prioritize remediation based on demonstrated impact. This approach may help reduce alert fatigue, a persistent challenge for teams managing large-scale application security programs.
After validation, CodeMender generates a proposed secure patch and presents it as a code difference for human review. Google explains that the agent uses an LLM-as-a-judge mechanism to assess whether the fix might disrupt existing functionality.
Organizations can also provide coding conventions and application-specific context to ensure that generated patches align more closely with their internal development standards. Developers retain final authority and must review and approve changes before they are committed.
CodeMender also integrates with Google AI Threat Defense and Wiz’s security platform. In this model, Wiz can enhance findings with deployment context from the Wiz Security Graph and invoke Red Agent for AI-driven penetration testing. At the same time, Green Agent coordinates remediation workflows and patch generation.
Google’s preview release positions CodeMender as part of a broader shift towards agentic, self-healing software development lifecycles.
Initially, organizations can access it using the generally available Gemini models through the Agent Platform, while CodeMender powered by Gemini 3.5 is limited to select governments and trusted partners.
ALERT: 20+ government sites delivered malware to businesses and citizens. See full attack research to check your own exposure.
