CodeMender
Google · CodeMender: Google Cloud's AI Agent That Finds, Verifies, and Fixes Code Vulnerabilities
CodeMender is Google Cloud's managed AI code-security agent, originally developed by Google DeepMind, that scans a codebase for vulnerabilities, verifies them in a local sandbox, and proposes tested patches for developers to review. It is in public preview for a limited set of Google Cloud customers and suits security and platform teams that already work in Google Cloud.
Best for
Security and platform teams already on Google Cloud who want an agent that verifies vulnerabilities and proposes tested patches, and who can join a preview program
Not ideal for
Teams that need a generally available product for production use today, organizations outside Google Cloud, and anyone looking for a free or self-serve tool
Who it's for
Enterprise security, AppSec, and platform engineering teams on Google Cloud
Capabilities
- Find, verify, and fix workflow: scan code, confirm findings, and generate tested patches
- Import findings from external scanners and have CodeMender verify and fix them
- Verification and patch testing run in a local process-level sandbox with outbound network blocked by default
- Patches are built and unit-tested, and by default you approve each file write; `cm fix` is kept from editing CI/CD, build, and repository configuration files outside the vulnerability's scope
- Local-first CLI (`cm`) that sends only targeted code snippets to the hosted agent
- Resumable sessions, reports in HTML, Markdown, JSON, or SARIF, a `cm health` pre-flight check, and pull-request delta scans with a CI failure threshold
- Choice of Gemini models per command, with Gemini 3.8 Flash as the default and Gemini 4 Argon, Google's frontier model, as an option
- Enterprise controls including VPC Service Controls, per-project isolation, and no training on customer source code
- Available on Gemini Enterprise Agent Platform or as part of AI Threat Defense with Wiz
- Documented CI setups for GitHub Actions (which can open a pull request with fixes for review), Cloud Build, and Git pre-commit hooks
Limitations
- Public preview for a limited set of customers; access is through Google Cloud sales
- Preview terms say it may be used only for limited testing and evaluation, not for commercial or production purposes
- Requires a Google Cloud project with the Vertex AI API, IAM permissions, and application default credentials
- You may analyze only code you own or are authorized to use, or open-source code under an OSI-approved license
- Google says the built-in sandbox is weaker than a fully isolated virtual machine, and Windows sandboxing is experimental
- Network access in the sandbox is all-or-nothing, so builds that fetch dependencies need pre-fetching or an open network
- The Gemini 3.8 Flash Cyber and Gemini 4 Argon Cyber Permissive models are restricted to Fairwind Program participants, and third-party model support was announced but not yet available
- Session data, including code snippets, is retained for up to seven days unless deleted
Use cases
- Scanning a service's source code and getting verified findings instead of a raw scanner list
- Importing results from an existing SAST tool and having CodeMender confirm and patch them
- Generating a tested patch for a confirmed vulnerability and reviewing the diff before committing
- Running delta scans on pull requests and failing CI on high-severity findings
- Connecting code-level remediation to cloud risk findings through AI Threat Defense and Wiz
Our take
CodeMender is aimed at the part of security work that usually stalls: turning a finding into a fix someone trusts. Verifying findings and testing patches against your build before proposing them is the right design, and the local-first CLI addresses code-privacy concerns. For now it is a preview for select Google Cloud customers, not something most teams can adopt for production, so treat it as one to evaluate rather than standardize on.
Who should use it
Google Cloud customers with security or platform teams who want to evaluate an agent that verifies vulnerabilities and drafts tested patches, and who can accept preview terms.
Who should skip it
Teams that need a production-ready, generally available tool today, organizations not on Google Cloud, and small teams looking for a free or self-serve option.
Strengths
- Verifies findings and tests patches before proposing them
- Local-first CLI sends only targeted snippets, with no training on customer code
- Imports findings from other scanners
- Token-based pricing at the underlying Gemini model's rate
- Backed by Google DeepMind research and Google Cloud enterprise controls
Weaknesses
- Preview only, for select customers, and not for production use yet
- Tied to Google Cloud and Gemini models for now
- Setup requires Google Cloud project, IAM, and CLI configuration
- Sandbox is weaker than full VM isolation
CodeMender pricing
Usage-based (Gemini 3.8 Flash)
$1.50 / $7.50 per 1M input / output tokens
- Introductory $0.75 / $3.75 through December 31, 2026, applied as 50% credits back on net spend
- Cached input $0.15 per 1M tokens
- Billed through Google Cloud
Usage-based (Gemini 4 Argon)
$4.00 / $20.00 per 1M input / output tokens
- Promotional 50% credits back on net spend through November 30, 2026
- Same list price for Gemini 4 Argon Cyber Permissive (Fairwind Program only), which the promotion does not cover
Usage-based (Gemini 3.1 Pro)
$2.00 / $12.00 per 1M input / output tokens
- Higher-capability model option
Note: Google states CodeMender pricing is token-based and determined by the model used; no separate seat fee is published. Other models in the CodeMender price table include Gemini 3.5 Flash at $1.50 / $9.00 and Gemini 3.8 Flash Cyber at $1.50 / $7.50 per 1M input / output tokens, with no introductory discount on the Cyber model. Access during preview is through Google Cloud sales. Verified on the Gemini Enterprise Agent Platform pricing page on 2026-10-23.
Technical specs
Available models
Where CodeMender excels
Reducing scanner backlog
Importing findings from existing tools and verifying them helps separate real issues from false positives before engineers spend time on fixes.
Drafting tested security patches
Patches are built and run against unit tests before review, so developers start from a candidate fix rather than a bare report.
PR-level security gating
Delta scans on pull requests and a CI failure threshold catch new issues without rescanning the whole codebase.
CodeMender vs. competitors
CodeMender vs. Codex Security
Codex Security is OpenAI's counterpart for the same find, reproduce, and patch loop, available in Codex products with an open-source CLI and SDK, GitHub pull-request reviews, and a cloud service for GitHub repositories. CodeMender is in a limited Google Cloud preview, runs through a local `cm` CLI with Gemini models, and bills per token through a Google Cloud project, so the choice largely follows whether your team works in OpenAI's or Google Cloud's ecosystem.
CodeMender vs. Google Mantis
Google Mantis is a separate open-source Google project that you deploy and modify yourself, with your choice of models and sandboxes, and Google labels it not officially supported and intended for demonstration rather than production. CodeMender is the managed Google Cloud product, in preview for select customers with token-based Gemini pricing, so it suits teams that want a supported path rather than a pipeline to own and customize.
CodeMender vs. Strix
Strix is open-source software that tests running applications and APIs and has its own cloud platform; CodeMender is a managed Google Cloud agent centered on verifying and patching vulnerabilities in source code.
CodeMender vs. Shannon
Shannon is a free, self-hosted white-box pentester that tests a running web app; CodeMender works on source code across many languages and focuses on producing tested patches for review.
Frequently asked questions
What is CodeMender?
CodeMender is an AI code-security agent from Google Cloud, originally developed by Google DeepMind. It finds vulnerabilities in a codebase, verifies them in a sandbox, and generates tested patches that developers review before committing.
Is CodeMender generally available?
No. CodeMender is in public preview for a limited set of customers under Google's Pre-GA terms, and access is through Google Cloud sales. The preview terms say it is for limited testing and evaluation, not commercial or production use.
How much does CodeMender cost?
Pricing is token-based and matches the Gemini model you use. For example, Gemini 3.8 Flash is listed at $1.50 per million input tokens and $7.50 per million output tokens, with introductory pricing of $0.75 and $3.75 through December 31, 2026, and Gemini 4 Argon is listed at $4.00 and $20.00. Google applies its promotions as 50% credits back on net spend. Usage appears in Cloud Billing.
Which languages does CodeMender support?
By default it covers C/C++, C#/.NET, Go, Java, JavaScript and TypeScript, Kotlin, Python, Ruby, Rust, and PHP, and other languages can be added in configuration. Google notes it does not publish per-language evaluations.
Does CodeMender change my code automatically?
Not by default. `cm fix` asks you to approve each file write unless you turn this off with `-y` or the `confirm_writes: false` setting. Google's GitHub Actions recipe can open a separate pull request with fixes for you to review and merge, rather than committing to your branch directly.
Integrations & fit
Alternatives to consider
OpenAI
Security and engineering teams that already use Codex or the OpenAI API and want an agent that validates findings and prepares reviewable fixes, from a one-off repository audit to per-PR checks in CI
Paid
Security engineers and AppSec teams who can operate isolated sandboxes and want a free, adaptable pipeline that reproduces and patches vulnerabilities with their own coding agent and models
Free
Strix (OmniSecure, Inc.)
Engineering and AppSec teams that want open-source, AI-driven pentesting of their own apps and APIs in the CLI or CI, with an option to move to a managed platform
Freemium
Keygraph
Engineering teams with access to both the source code and a staging copy of their web app or API who want open-source, evidence-backed security testing in CI
Freemium
Anthropic
Security and engineering teams at Claude Enterprise organizations that want continuously updated, validated findings across their own repositories, with severity tuned to their threat model and fixes handled in Claude Code
PaidAbout CodeMender
CodeMender splits work between a hosted multi-agent system and a local `cm` CLI that also acts as a daemon on your machine. You run `cm find` on a path to scan for vulnerabilities, or import findings from other scanners; `cm verify` builds the code and checks whether a finding is real inside a local process-level sandbox; and `cm fix` generates a patch, applies it to a copy of the code, runs your build and unit tests, and re-checks the finding before asking you to approve each file write (auto-approval is opt-in). Full repositories are not uploaded; the CLI sends targeted snippets and tool results, and Google says customer source code is not used to train models. Findings and patches are exported as HTML, Markdown, JSON, or SARIF, and recent CLI releases added pull-request delta scans, a CI gate, and deep scanning by default, with documented recipes for GitHub Actions, Cloud Build, and pre-commit hooks. You access it through Gemini Enterprise Agent Platform with generally available Gemini models (Gemini 3.8 Flash by default, with Gemini 4 Argon as an option since October 13, 2026), or as part of Google's AI Threat Defense alongside Wiz; the specialized Gemini 3.8 Flash Cyber and Gemini 4 Argon Cyber Permissive models are limited to organizations approved for Google's Fairwind Program. Billing is token-based at the price of the chosen model. The tradeoffs: it is pre-GA, available only through sales to select customers and not licensed for commercial or production use during preview, it requires a Google Cloud project, and the built-in sandbox is weaker than a fully isolated VM.
Updates from CodeMender
CLI v0.16.0 made `cm find` run a file-by-file deep scan by default (`--quick` keeps the single-session scan), added a `cm health` command that checks credentials, configuration, and project authorization before a scan, and widened file discovery to more C++ headers and shebang scripts.
CodeMender added Gemini 4 Argon as a model option for `cm find`, `cm verify`, and `cm fix`. Gemini 3.8 Flash remains the default, and Gemini 4 Argon Cyber Permissive is available to Fairwind Program customers.
CLI v0.10.0 added `--diff` and `--staged` scans of changed code, SARIF v2.1.0 export, a `--fail-on` severity gate for CI, and deep scanning with `--deep`.
A September CLI release expanded the default language set to C/C++, C#/.NET, Go, Java, JavaScript and TypeScript, Kotlin, Python, Ruby, Rust, and PHP.
CodeMender switched its default model to Gemini 3.8 Flash, and Gemini 3.8 Flash Cyber became available to allowlisted Fairwind Program customers.
Google Cloud made CodeMender available in public preview for select customers through Gemini Enterprise Agent Platform and as a component of AI Threat Defense.
Are you the founder? Claim this listing →