Add preferred source
Google has enhanced CodeMender, its AI agent that finds and fixes software vulnerabilities. Leveraging multiple AI models including Gemini 4 Argon, the tool now lets users choose between a sandbox or a virtual machine as the verification environment, and narrow the scope of investigation to specific folders or only changed files. Local operations incorporate human confirmation. Google Cloud also announced the general availability of Auth Manager, which manages AI agent credentials, and stronger threat detection through its partnership with Wiz.
Key Elements
Google announced it has enhanced CodeMender, its AI agent that discovers and fixes software vulnerabilities. The company disclosed the update in an official blog post dated November 4, expanding the options for the isolated environment used to verify vulnerabilities and allowing users to specify the scope of files to be investigated. The tool is currently offered as a preview.
CodeMender runs by combining multiple AI models, including Gemini 4 Argon. It is said to have the ability to identify previously undiscovered vulnerabilities and generate verified fixes before attackers can exploit them. In its announcement on the 5th (US time), Google Cloud also cited the tool’s new features as part of its broader effort to strengthen security capabilities for the AI agent era.
With this enhancement, the environment used to verify vulnerabilities and reproduce attacks can now be selected from two types: a sandbox that restricts program behavior using OS features, or a virtual machine that virtually constructs an independent computer. Running verification in an isolated environment reduces the risk of problems arising during work. In addition, a mechanism has been added that requires human confirmation for file writes and command execution in the local environment.
For investigation scope, users can choose from three options: a specific folder, the entire repository, or only files changed since the previous investigation. By narrowing the target—for example, examining only the changes—the time and cost required for investigation can be reduced.
| Item | Before enhancement | After enhancement |
|---|---|---|
| Verification environment | Reproduce attacks in an isolated environment | Choose sandbox or virtual machine |
| Local operations | No specific provisions | Human confirmation for file writes and command execution |
| Investigation scope | Cannot be specified | Folder / entire repository / changed files only |
| Applying fixes | Requires developer review and approval | Same as before (unchanged) |
Note: Compiled based on Google’s official blog and Google Cloud’s announcement.
Since the preview launch in July, Google has explained that attacks can be reproduced in an isolated environment managed by the user. The main pillars this time are making it possible to choose the type of isolated environment and supporting the specification of investigation scope. Meanwhile, the company continues to require developer review and approval before AI-generated fixes are applied to a repository.
Beyond CodeMender, Google Cloud has rolled out a series of features aimed at governing AI agents. Auth Manager, which centrally manages AI agent credentials, has entered general availability. It handles credentials in storage managed by Google and automates authentication flows and token refresh
With “Agent Identity,” the company is offering a preview of support for JWT (JSON Web Token). By issuing standard OIDC tokens, it enables AI agents to authenticate to AWS, Azure, on-premises environments, MCP servers and others without using long-lived static API keys.
For threat detection and response, Google is deepening its partnership with security company Wiz. “Wiz AI SAST,” currently in preview, uses AI to analyze application behavior and detect vulnerabilities hidden in complex business logic, and also supports remediation in coordination with tools such as CodeMender.
Integration between “Wiz Defend” and “Google Security Operations” has also been improved. By adopting a common data model, threats can be investigated from either product, and investigation results can be synchronized bidirectionally. Wiz Defend is also integrated with “Google Threat Intelligence,” allowing the latest threat information to be ingested directly.
Behind this series of announcements is the spread of agentic coding tools. While development speed has increased dramatically, human review cannot keep up with commits made at machine speed or with unverified dependencies, creating room for attackers to plant backdoors. Google Cloud says it aims to build real-time guardrails into integrated development environments (IDEs), validating models and eliminating defects before code is merged.
The security of the software supply chain is becoming a management issue for corporate development teams. How quickly the time from vulnerability discovery to remediation can be shortened is also a factor that influences the competitiveness of cloud providers. If automated remediation tools become more accurate, it could help ease the burden on companies struggling with a shortage of security personnel. However, the governance question of who ultimately confirms the validity of AI-generated fixes remains, and companies will continue to face scrutiny over how they design their operations.
