CodeMD Call-Path Security · Our own testing on public advisories

What we found replaying real advisories

We took published command-injection vulnerabilities in open-source MCP servers, went back to the commit that added each bug, and ran the CodeMD Call-Path Security check on it, exactly as it checks a pull request.

12 of 14failed the check on the commit that added the bug
+2flagged in the repository's first commit
2–15 monthseach bug was in public code before its advisory

These are results from our own test, with a link to every public advisory and commit so you can check them yourself. They show what the check found in these 14 cases; they aren't a guarantee it will find every issue in your code.

Blocked on the pull request that added the bug

In each one, a tool an AI agent can call passes its input to a shell command. All are fixed now; this isn't a criticism of the maintainers. Links go to the public advisory and the commit the check would have failed.

mcp-nmap-server

GHSA-xc68-rrqc-qgq3 · CVE-2026-3484 · command injection
✗ Would have failed
Commit it fails on
"Add files via upload", January 9, 2025
Path the check reports
MCP tool call from an AI agent → runNmapScan: process/shell execution
Advisory published
March 3, 2026, about 14 months later

mcp-maigret

GHSA-2g7v-hghf-grg4 · CVE-2026-2130 · command injection
✗ Would have failed
Commit it fails on
"Migrate to Docker Method", December 13, 2024
Path the check reports
MCP tool call from an AI agent → MaigretServer.execCommand: process/shell execution
Advisory published
February 8, 2026, about 14 months later

mcp-markdownify-server

GHSA-45qj-4xq3-3c45 · CVE-2025-58358 · command injection in the pptx-to-markdown tool
✗ Would have failed
Commit it fails on
"feat: turn into generic markdownify module", December 20, 2024
Path the check reports
MCP tool call from an AI agent → createServer → Markdownify.toMarkdown → Markdownify._markitdown: process/shell execution
Advisory published
September 2, 2025, about 8 months later

@elgentos/magento2-dev-mcp

GHSA-xqv9-qr76-hfq2 · CVE-2026-5603 · command injection
✗ Would have failed
Commit it fails on
"Added cache commands", September 28, 2025
Path the check reports
MCP tool call from an AI agent → executeMagerun2Command: process/shell execution
Advisory published
April 6, 2026, about 6 months later

mcp-package-docs

GHSA-vf9j-h32g-2764 · CVE-2025-54073 · command injection in several tools
✗ Would have failed
Commit it fails on
"fix: docs parsing for npm (#8)", March 5, 2025
Path the check reports
MCP tool call from an AI agent → PackageDocsServer.searchPackageDocs, plus 5 more functions that run local Go and Python tooling: process/shell execution
Advisory published
August 5, 2025, about 5 months later

mcp-for-stata

GHSA-4p62-hqp5-g644 · CVE-2026-47708 · command injection through the log file name
✗ Would have failed
Commit it fails on
"refactor(stata)!: promote stata to top-level module", March 17, 2026. This commit moved the code into a new module, so the check saw the risky path as new.
Path the check reports
MCP tool call from an AI agent → StataDo._execute_windows, plus 1 more function: process/shell execution
Advisory published
June 4, 2026, about 3 months later

Flagged in the first commit

These bugs were in the repository from its first commit. The check flags the path, but a repository's first check only records a starting point, so today it would show ✓. Installing the app before writing the code, or a full scan with the Security panel on Analyze Repo, would show them.

adb-mcp

GHSA-54j7-grvr-9xwg · CVE-2025-59834 · command injection
First commit
Commit
"Initial commit", April 12, 2025
Path the check reports
MCP tool call from an AI agent → executeAdbCommand: process/shell execution
Advisory published
September 24, 2025, about 5 months later

mcp-server-semgrep

GHSA-86hp-qxqp-w9wv · CVE-2026-7446 · command injection
First commit
Commit
"Initial commit", February 1, 2025
Path the check reports
MCP tool call from an AI agent → SemgrepServer.handleScanDirectory, plus 7 more handlers: process/shell execution
Advisory published
April 30, 2026, about 15 months later

Real bugs, found and fixed before they cost anyone

The same review CodeMD runs on your pull requests, run on real products. It found real bugs, not style nits, across security, correctness, reliability and data integrity. Every one was fixed.

33real bugs found and fixed
11security bugs among them
24 of 33now guarded by a test so they stay fixed
CategoryWhat it protects
SecurityYour code and data stay yours.
CorrectnessResults you can trust.
ReliabilityWork that finishes, every time.
Data integrityNothing you saved is lost or overwritten.

Counts are our own findings, not an outside audit.

When changed-file detection becomes the failure

A public Sentry workflow shows a different problem CodeMD is designed to avoid: a large change produced so much repeated file-list data that the tool deciding what should run could not start.

Sentry meta(deploy)

Public GitHub Actions run 42934913116 · change detection failed before its JavaScript ran
Observed in public CI
What happened
The workflow expanded about 239,000 characters into one actions/github-script input. Overlapping categories repeated hundreds of paths: 418 testable files, 421 type-checkable files, 429 frontend-lint files and 704 acceptance files.
Failure
Argument list too long. The runner could not start Node, so the script never classified the change.
Why CodeMD
A list of changed files is not a test plan. CodeMD maps changed functions through the call graph to the tests that reach them, then hands CI a bounded set of impacted tests instead of copying the full path list into a process argument.
Verify it
See the run and error and the public workflow source.

CodeMD did not run on this repository. This is an independently observed public CI failure that demonstrates the scaling problem, not a claim that CodeMD found or fixed Sentry's workflow.

When changed-file detection becomes the failure

A public Sentry workflow shows a different problem CodeMD is designed to avoid: a large change produced so much repeated file-list data that the tool deciding what should run could not start.

Sentry meta(deploy)

Public GitHub Actions run 42934913116 · change detection failed before its JavaScript ran
Observed in public CI
What happened
The workflow expanded about 239,000 characters into one actions/github-script input. Overlapping categories repeated hundreds of paths: 418 testable files, 421 type-checkable files, 429 frontend-lint files and 704 acceptance files.
Failure
Argument list too long. The runner could not start Node, so the script never classified the change.
Why CodeMD
A list of changed files is not a test plan. CodeMD maps changed functions through the call graph to the tests that reach them, then hands CI a bounded set of impacted tests instead of copying the full path list into a process argument.
Verify it
See the run and error and the public workflow source.

CodeMD did not run on this repository. This is an independently observed public CI failure that demonstrates the scaling problem, not a claim that CodeMD found or fixed Sentry's workflow.

How we tested

  1. Picked the advisories: published command-injection advisories (CWE-78) for open-source MCP servers in GitHub's Advisory Database that link to the commit that fixed the bug.
  2. Found the commit that added each bug: we walked every earlier commit that touched the fixed files, oldest first, and checked each one against the commit before it, the way the GitHub check checks a pull request.
  3. Counted it as blocked only if the check failed on a function that was still risky in the version the fix changed. The first such commit is the one listed.
  4. Used the same free check the GitHub App runs: no AI and no hand-tuning for these projects. General fixes to the check came from this test and apply to every repository: recognising execAsync = promisify(exec), tool handlers registered outside a function, typed server.tool<…>( registrations, and calls inside nested callbacks.

A failed check shows a path from outside input to a shell command. It doesn't prove the command is built unsafely: a reviewer decides that, with the AI verdict that CodeMD Pro adds to the check.

Catch the next one before it ships.

Free for public repositories. One click to install.

Install free on GitHub