auth_* rules skip because
DeepWiki serves anonymous requests, so there is no authorization flow to
grade. A server behind a login grades them; see
authenticated servers.
The mapping
A risk is covered in part when mcpscore checks some of what the risk describes. No risk is covered in full: each one also has a part that only code, configuration or the client can show.
One Security & Auth rule maps to no OWASP risk.
security_malformed_request_handling checks that a malformed request gets the
JSON-RPC parse error,
which is error hygiene rather than one of the ten risks.
The specification sets the same footing for these checks. Its
§Security and Trust & Safety
says tool descriptions are untrusted unless they come from a trusted server,
and that implementers should follow security best practices. The Origin and
TLS rules enforce requirements the spec states directly, in
Transports §Security Warning
and Authorization.
Every rule’s citation is in the rules reference.
Where an audit from outside stops
An audit sees what a server publishes and how it answers requests. It never calls a tool, never reads source code and never sees the client. Each gap below follows from one of those limits.- MCP01: mcpscore finds a credential the server publishes. It does not see how the server stores tokens, how long they live, or what it writes to its logs.
- MCP02: mcpscore sees that scopes are advertised. What a granted token can actually do is decided inside the authorization server and the tools.
- MCP03 and MCP06: the catalog-text rules read for hidden characters and a fixed list of phrasings. A description that misleads in plain words passes. Text returned by a tool is out of reach, because the audit never calls one. A catalog that changes after the audit is out of reach too; running mcpscore in CI catches the change on the next run.
- MCP04: a package audit reads registry metadata only. It does not download, run or scan the package or its dependencies.
- MCP05: injection happens when a tool runs with untrusted input. The
audit never sends
tools/call. Smoke mode calls your own read-only tools with inputs derived from their schemas, not attack payloads. The exceptions are opt-in: the filesystem boundary check sends file tools paths that try to escape their root, and the network boundary check sends fetch tools URLs that try to reach a local listener their policy denies. - MCP07: mcpscore checks the authorization discovery chain and that a made-up bearer token is refused. Whether each tool enforces authorization, and whether the server rejects a real token issued for another service, needs credentials and tool calls.
- MCP08: logging and telemetry live inside the server.
- MCP09: an inventory of the MCP servers running in an organization is a governance task. mcpscore audits the server you point it at.
- MCP10: context is shared or kept in the client and in server-side session state, neither of which a request from outside can observe.
Local servers
When you audit a server onlocalhost, 127.0.0.1 or ::1, two more rules
run. They skip for any other address, so they never run in the web audit.
security_host_header_validation(MEDIUM): DNS rebinding points a website’s hostname at 127.0.0.1, and the browser then sends that hostname asHost. The rule sends a request the server accepts, then the same request withHost: mcpscore.invalid, and passes when the second one is refused with any 4xx. The official Python and TypeScript SDKs do this by default for a server bound to localhost.Originvalidation stays the main defence; this is the second layer.security_local_bind_loopback(MEDIUM): the transport spec says a server running locally should bind only to localhost, not to all network interfaces. The rule connects to the same port on this machine’s network addresses and sends the request the server accepted on localhost. It fails when the server answers there: it listens on0.0.0.0or::, and other machines on the network can reach it. Another program on that port does not count, and a connection that times out leaves the rule unjudged. It runs forhttp://targets only. The report records how many addresses were tried, never the addresses.
How the catalog-text rules decide
The five catalog rules read every string a server publishes: instructions, server info, names, titles, descriptions, schema strings, URIs, MIME types,_meta and icon URLs. A finding names where the string is and what class
matched. It never repeats the string, so a report cannot leak a credential or
replay an injected instruction.
Technical details
Technical details
catalog_hidden_unicodeflags Unicode tag characters outside a valid emoji tag sequence, bidirectional overrides and isolates, control characters other than tab and line breaks, and runs of two or more zero-width characters, directional marks, soft hyphens or variation selectors. One such character alone passes, because emoji and right-to-left text use them.catalog_no_embedded_secretsmatches credentials by their provider’s issued format: AWS access key IDs, GitHub, GitLab and Slack tokens, Stripe live keys, Anthropic, OpenAI and Google API keys, private-key headers, JWTs and long bearer tokens. Documentation samples and placeholder values pass.catalog_prompt_injection_phrasingmatches a fixed list of directives to the model: overriding earlier instructions, keeping an action from the user, reassigning the model to a persona, and chat-template control tokens. A phrase in matched quotes, listed with a slash, or introduced as an attempt (“attempts to override the system prompt”) names an attack rather than performing it, and passes. No model judges the text.catalog_exfiltration_linksreads each string as Markdown, the way a client renders it, including reference-style links and raw HTML. In HTML it reads the attributes a browser fetches on render (<img src>,srcset,<video poster>,<object data>, a stylesheet, icon or preload<link href>, SVG<image href>, a refresh<meta>, a<script src>that runs (a data block such as JSON-LD does not),url()and the strings@import,image-set()and custom properties take, in astyleattribute on any element or in a<style>block,url()in an SVG presentation attribute such asfillorcursor, and the values an SVG<set>or<animate>gives anhrefor such an attribute) or follows on a click (<a href>,<form action>); the same attribute on an element that ignores it is inert. Adns-prefetchorpreconnect<link>sends only its hostname, so only the hostname is checked. It flagsjavascript:andvbscript:links,data:links other than a raster image, and a link or image URL with a slot for the model to fill, such as{conversation},{{secret}},${TOKEN}or<data>. An image loads without a click, so a slot anywhere in its URL counts; a link counts only with the slot in its query, because a path such as/users/{id}is usually an API route. A form action’s query counts whatever the form’s method. Code spans and image alt text are text, not links, and pass, and so does the content of a<template>(unless it declares a shadow root),<script>and<textarea>, and of a<noscript>except in an iframesrcdocsandboxed withoutallow-scripts, where it renders. A relative URL is resolved against the document’s<base href>. Markup that cannot be checked fails asuncheckable_markup: a string with markup over 64 KiB (real catalog strings stay far below), Markdown whose reference links, or a long<base href>, would repeat a destination past four times that size, iframesrcdocdocuments nested more than three deep, and constructs that browsers and parsers read differently: a<base href>whose host is not a plain ASCII domain, IPv4 or IPv6 address, a CDATA section in HTML, a comment closed early by<!-->or--!>, a<frameset>, a raw-text tag,<base>or SVG/MathML inside a<select>, an end tag that would close a<select>, raw-text content such as<script>or<xmp>inside SVG or MathML, SVG or MathML inside a<template>,</form>,</body>or</html>with elements still open inside, an end tag that closes unclosed elements across the edge of SVG or MathML content, out of a shadow-root<template>, past an open block such as<div>for a custom element, or out of a<video>,<audio>or<picture>, a<source>inside a media element but not its direct child, or a raw-text end tag such as</script/>that a browser accepts and the parser reads as text.catalog_cross_tool_directivesreads each sentence that directs the model: one withmust,should,make sureorensure;always,neverordo notopening a clause (“Never use …”, ”…, then always call …”, not “the tool always returns”);firstorinsteadbefore a verb such ascall; one that opens with a verb such ascalloruse, also afterplease,thenornext; or one that opens withwhen,before,afterorifand whose main clause does not start with a subject (“When asked, call …”, not “When available, the tool returns …”). Cues inside quotes or inline code do not count. In such a sentence it flags a reference to any or other tools or servers (“with any tool”, “using another server”, “every tool must …”) asany_tool_directive, and a tool name the server does not list asunlisted_tool_directive. A tool name is code-shaped: in backticks and made of tool-name characters (letters, digits,_,-,.and/, so notiam:CreateRole), joined with_or., or camelCase, or a/path with such a segment (mail/send_email, not “read/write”), aftercallorinvokeor before the wordtool. Names of the server’s own tools, prompts and resources pass, also behind a client namespace such asmcp__notes__searchornotes.search, and so do plain-word mentions such as “pairs well with the GitHub server” and phrases in double quotes or named as attempts (“attempts to call …”). Unlisted names are judged only when every catalog the server declares was listed whole, from the same lifecycle as the text (without a session, the stateless tools page must sayresultType: "complete"). MCP method names such asresources/readandpingare protocol, not tools, and pass.- All five skip a partial audit as
insufficient-data, because a server behind a login shows no catalog without credentials. skills_hidden_unicode,skills_no_embedded_secretsandskills_prompt_injection_phrasingrun the same checks on a server’s skills: every frontmatter string and every line of eachSKILL.mdread. A finding names the skill and the line. Chat-template tokens inside a code block or an inline code span count as an example, not a directive; every other class counts inside code too.
mcpscore.toml. The public score on mcpscore.dev keeps
every rule.
Security findings in CI
--sarif writes every failed rule as a code scanning alert. Security & Auth
rules carry GitHub’s security-severity, so they sort into the Security tab’s
critical, high, medium and low bands next to your other security alerts.
security_origin_validation finding has level error and
security-severity 7.0, which code scanning files as high.
The CLI reference has the full mapping.
The GitHub Action writes the
same file through its sarif-path input.
What’s next
- Scoring methodology: how Security & Auth counts in the score.
- Authenticated servers: grade the
auth_*rules on a server behind a login. - Rules reference: every Security & Auth rule with its severity and the spec revisions it applies to.
- The OWASP MCP Top 10, and what your server’s catalog tells a model: the catalog-text rules on a demo server, and how they were measured on live servers before release.
- The LiteLLM MCP auth bypass, and the one request that shows it:
CVE-2026-59822 and how
auth_invalid_token_rejectedsends a made-up token.