NewApp Monitoring is liveSign up free
AgentbeamBeta
AI agent vulnerability · Other MCP servers and tools

CVE-2026-11720: MCP Toolbox HTTP tool path parameters allow traversal to unintended endpoints

criticalCVSS 4.0 9.3 · 3.1 9.1GHSA-vwxw-jrg6-9jxv

MCP Toolbox (`github.com/googleapis/mcp-toolbox`) before 1.3.0 lets a client escape the path configured for an HTTP tool. Path parameters containing `../` are normalized away when the URL is resolved, so the toolbox calls other endpoints on the same host and sends its own credentials. Upgrade to 1.3.0.

Published Jun 29, 2026 · Updated Sep 4, 2026 · Analysis Nov 1, 2026

What an attacker can do

An HTTP tool in MCP Toolbox has a configured path template such as /api/v1/users/{{.id}}. The URL builder substitutes the caller's path parameters into that template and parses the result as a relative URL. It checks that the scheme, host and user info did not change. It then calls ResolveReference, which normalizes dot segments. A value like ../../admin/secrets therefore climbs out of the intended scope.

The request goes to the same host, but to an endpoint the operator never exposed, and the toolbox attaches the credentials it was configured with. The attacker is whoever can call the tool: an MCP client, or an AI agent steered by a prompt injection that fills in the parameter. They need no authentication to the downstream API themselves. The CVSS vector is network, no privileges, no user interaction, high confidentiality and integrity. In practice a tool built to read one user record can be turned into a way to read or modify any endpoint the toolbox's credential can reach.

Who is affected

All versions of github.com/googleapis/mcp-toolbox below 1.3.0 are affected. You are exposed if you define HTTP tools that insert parameters into the URL path. Check the version of the binary or module you run.

go list -m github.com/googleapis/mcp-toolbox

If you run a release binary or container, check its version flag or image tag and compare it with 1.3.0. Then review your tools config for path values containing {{.param}} placeholders.

github.com/googleapis/mcp-toolbox

Ecosystem
Go
Affected
< 1.3.0
Fixed in
1.3.0
Check
go list -m github.com/googleapis/mcp-toolbox

How to fix it

Upgrade to MCP Toolbox 1.3.0.

go get github.com/googleapis/[email protected]

The fix (pull request 3218) rejects relative path segments (..) and their encoded form (%2e%2e). It also requires the resolved URL to stay inside the configured base path, and it blocks sibling paths that only share a string prefix. It adds pathEscape and queryEscape template functions so you can encode parameter values safely when you insert them into paths or query strings. The sources give no other workaround. Until you upgrade, limit the downstream credential to the minimum scope.

How to tell if you were hit

Check the downstream API's access logs for requests from the toolbox's credential to paths outside those your HTTP tools define, especially paths containing no .. because they were already normalized, such as admin or secrets endpoints. Compare each logged path with the tool's configured template. Review MCP tool-call logs for parameter values containing ../ or %2e%2e. If an AI agent drives the toolbox, Beam records MCP tool calls, and its rules can flag the ones with traversal sequences in their arguments.

Original advisory

From the GitHub Advisory Database (CC BY 4.0).

A path traversal vulnerability exists in the HTTP tool URL builder of googleapis/mcp-toolbox.

When constructing downstream API requests, the URL builder substitutes user-controlled pathParams into the configured tool path and parses the resulting string as a relative URL. While it checks that the input does not alter the scheme, host, or user info, it relies on ResolveReference for the final URL resolution. Because dot segments (../) are normalized during this resolution step, an attacker can supply path parameters containing directory traversal sequences to escape the operator-configured path scope. This allows the client to coerce the toolbox into making requests to unintended endpoints on the same target host while forwarding the toolbox's configured credentials (e.g., bypassing a restricted path like /api/v1/users/{{.id}} to reach /admin/secrets).

Questions

Does CVE-2026-11720 need authentication to the toolbox?

The CVSS vector says no privileges are required. Anyone able to invoke the HTTP tool, including an AI agent acting on injected instructions, can supply the malicious path parameter.

Which MCP Toolbox version fixes it?

1.3.0. Every version before it is affected.

Can the attacker change the host that the toolbox calls?

No. The builder still checks that the scheme, host and user info are unchanged. The flaw is limited to other paths on the same host, but those requests carry the toolbox's credentials.

References

How this page is made: the data panel comes from the GitHub Advisory Database, CISA’s KEV catalog, the NVD and FIRST EPSS, and updates twice a day. The analysis is written by an AI model from the sources listed above and checked by automated rules before it is published. Check versions against the vendor's advisory before you act. Methodology

More Other MCP servers and tools vulnerabilities

All Other MCP servers and tools entries