MCP Inspector could run code from a remote server’s OAuth URL
Short answer
On October 21, 2026, the MCP Inspector project published GHSA-q978-78c7-phqw. Versions 2.0.0 through 2.11.0 of @modelcontextprotocol/inspector could run arbitrary code as the user on Windows and WSL while connecting to a remote MCP server. The server chooses an OAuth authorization URL, and the npm open package placed that URL in a PowerShell double-quoted string, which evaluates subexpressions. No tool call and no model decision are required. macOS and Linux are not affected. The patched release is 2.12.0. The advisory rates the issue high.
Last updated · Sekrd Research
- Advisory: GHSA-q978-78c7-phqw, published October 21, 2026 by cliffhall. High. CWE-78. Reported by eilonc-pillar.
- Package: @modelcontextprotocol/inspector 2.0.0 through 2.11.0. Patched in 2.12.0.
- Where: Windows and WSL. macOS and Linux use an argument vector, not a shell, and are not affected.
- Trigger: connecting to a remote MCP server. The server returns 401 and names the authorization URL. Auto-open is on unless MCP_AUTO_OPEN_ENABLED is false.
- CVSS vector: CVSS:3.1/AV:N/AC:L/PR:N/UI:R/S:U/C:H/I:H/A:H. The advisory’s impact table says no click and no model choice are required.
What happened
MCP Inspector is the debugging client people point at an MCP server they do not yet trust. On October 21, 2026, the project disclosed that a remote server could run code on the developer’s Windows or WSL machine during the connection itself. The model is not involved. No tool is called. Nothing has to be approved. The advisory’s impact table says a click is not required.
The server answers 401, serves OAuth discovery documents, and puts whatever string it wants in authorization_endpoint. Inspector opens that URL automatically. On Windows, and on WSL outside a container, the dependency open at 10.2.0 places the URL inside a PowerShell double-quoted string and runs it. PowerShell evaluates $(...) subexpressions inside double quotes. URL serialization keeps the characters that payload needs. The advisory shows that the application path in the same function was escaped, and the URL was not.
The CVSS vector printed on the advisory includes UI:R, which usually means a user has to participate. The same page’s impact table says “Requires a click or approval: no.” Both statements are on the advisory. This article does not pick one and drop the other.
Timeline
- October 21, 2026. GHSA-q978-78c7-phqw published. Affected: 2.0.0 through 2.11.0. Patched: 2.12.0.
- Dependency detail: the three package.json files pin
opento^10.2.0. Version 11.0.0 escapes the PowerShell argument. A caret on 10.x does not install 11.x, so a routine npm update inside the 10 line does not pick up that fix. Inspector 2.12.0 is the release that moves the caller. - Not this bug: GHSA-28xh-wpgr-7fm8 for
opencovers versions below 6.0.0. The advisory says the PowerShell branch described here is in 10.2.0 and was fixed separately in 11.0.0.
Why connecting is enough
Inspector’s job is to attach to a server whose behavior is unknown. The advisory says that is the population most exposed, by design. The payload is not stored in a config file you can review on clone. It arrives in the discovery response at request time. A registry entry, a typosquat, an expired domain, or a server that was fine when you added it and compromised later all produce the same result, because the URL is chosen when you connect.
Auto-open is controlled by MCP_AUTO_OPEN_ENABLED. The value true forces it on, false forces it off, and the default is on unless Vitest is running. The remaining condition the advisory names is that stderr is a TTY, which it is when a developer runs Inspector in a terminal. There is no prompt.
The same open call exists in the CLI, the TUI, and the web server config. macOS and Linux are outside the bug because open and xdg-open are spawned with an argument list. The advisory does not claim a sandbox escape or a privilege escalation past the user who launched Inspector. The reporter’s executed proof of the PowerShell string was run on macOS PowerShell, with the Windows and WSL outcome argued from the platform branch plus that proof. The advisory says that limit plainly.
A related class of bug, CVE-2025-6514, was fixed in mcp-remote by percent-encoding $ before the same kind of open. The Inspector advisory lists that approach, moving to open 11, or checking that the authorization URL is https (or http on loopback) and on the same origin as the configured server. open’s own readme says the package makes no security guarantee if you pass untrusted input.
How to check if your app is exposed
- On any Windows or WSL machine where you debug MCP servers, check the installed Inspector:
npm ls -g @modelcontextprotocol/inspectorand the copy inside the repo. Versions 2.0.0 through 2.11.0 are affected. Install 2.12.0. A floating range that stays on open 10.x will not cross the major bump by itself. - Until that install is in place, set
MCP_AUTO_OPEN_ENABLED=falsein the environment you use to launch Inspector, so it does not open the authorization URL for you. - Treat a remote MCP server like a program you chose to run. If you connected Inspector to a server you do not operate, from Windows or WSL, on an affected version, rotate credentials that lived on that user account and review what the server was. The advisory does not publish a list of malicious servers.
- For an MCP server you ship, require authentication on anything reachable from the internet, and do not treat a successful unauthenticated
tools/listas healthy. That is a different check, described in the MCP checklist. - Read
.cursor,.vscode, and repo MCP config before you connect a coding agent. This Inspector bug is in the client, not in those files, but the files decide which server the client will call.
How Sekrd helps
Sekrd does not scan the Inspector binary on a developer laptop. A deep scan of a deployed app calls tools/list without credentials when it finds an MCP endpoint, and reports a successful list. It does not call any tool. The rule is mcp-unauth-tools. If you point a deep scan at a GitHub repo or a config file, the MCP review looks for tool poisoning, an over-permissioned filesystem, hardcoded secrets, and command injection in that source. None of that replaces upgrading Inspector.
Cursor-built apps are the ones most likely to grow an MCP server during a normal session. The landing page is /cursor-security-scanner. After you patch Inspector, scan the app you actually deployed: free scan for the bundle, deep scan if an MCP route is part of the product.
Related reading
Frequently asked questions
Which MCP Inspector versions are affected?
2.0.0 through 2.11.0 of @modelcontextprotocol/inspector. The patched version is 2.12.0. The advisory was published on October 21, 2026.
Are Mac and Linux affected?
The advisory says no. On those systems open uses an argument vector. Windows and WSL are the affected platforms.
Does the model have to call a tool?
No. The advisory says the chain runs while connecting, with no tool call and no model involvement. The impact table also says no click is required. Auto-open is the default unless MCP_AUTO_OPEN_ENABLED is false.
Will npm update inside my project pick up the fix?
Not if the project only floats open on the 10.x line. The advisory says a caret range on 10.2.0 never resolves to open 11. Install Inspector 2.12.0.
Is this CVE-2025-6514?
No. The advisory cites CVE-2025-6514 as the earlier mcp-remote bug that was fixed by sanitizing the same class of URL. GHSA-q978-78c7-phqw is the Inspector issue fixed in 2.12.0.
Sources
Don't ship until you're sekrd
Run a free scan to find the vulnerabilities your AI missed.
Scan Your App Free