MCP server vulnerabilities by flaw type

What the 267 known vulnerabilities in MCP-named packages actually are, by the flaw type their own advisories give, and the question each type puts to a review. Records as of November 13, 2026.

Drel10 min read

Published by Drel · Method · Limits · Download the data (CSV) · Records as of November 13, 2026

MCP is new. The flaws reported in MCP servers mostly are not. This page takes every known vulnerability in a package whose name says MCP and groups it by the weakness type (CWE) its own advisory assigns. The grouping is done by a fixed table, in code. Nobody at Drel found, tested or reviewed the individual vulnerabilities, and no model classified them.

Findings

  • As of November 13, 2026, public advisories describe 267 known vulnerabilities in MCP-named packages, and 260 of them (97%) carry a weakness type assigned by the advisory itself.
  • 147 of those 260 (57%) are in at least one of three long-known classes: command and code injection; path traversal and file access; and server-side request forgery.
  • Command and code injection accounts for 61 of the 260 classified MCP vulnerabilities (23%).
  • Path traversal and file access accounts for 51 of the 260 classified MCP vulnerabilities (20%).
  • Server-side request forgery accounts for 40 of the 260 classified MCP vulnerabilities (15%).
  • 167 of the 267 (63%) are rated critical or high by their own advisory.
  • The first was published on May 29, 2025. 94 were published in the third quarter of 2026 alone.
  • 24 of the 267 are in packages the Model Context Protocol project publishes itself. The other 243 are in servers and tools built by others.
  • The single most common weakness identifier is CWE-22, on 46 vulnerabilities, followed by CWE-918 on 40.

By flaw type

“Critical or high” is the share of that type's vulnerabilities their own advisories rate critical or high. The last column is the type's share among the model SDKs and agent frameworks on Drel's lists (183 classified vulnerabilities), for comparison.

Flaw typeMCP vulnerabilitiesShare of MCPCritical or highShare in SDKs and frameworks
Command and code injection6123%66%20%
Path traversal and file access5120%69%8%
Server-side request forgery4015%53%11%
Missing or weak authentication3815%82%7%
Missing or wrong authorization3513%69%9%
Information exposure249%33%3%
Origin checks and cross-site flaws218%76%4%
Denial of service145%71%13%
SQL and query injection42%50%6%
Unsafe deserialization00%n/a8%
Other weakness types3212%26%

What each type looks like, and what to check

For each type: how it shows up in an MCP server, the question it puts to a review, and up to three public advisories of that type, the most severe first. The advisory titles are the advisories' own.

Command and code injection · 61 (23%)

A tool builds a shell command, a script or a template from its arguments. The arguments come from the model, and the model repeats whatever text it was given.

Check: For every tool that runs a process or evaluates code: are arguments passed as a list, never as a string to a shell, and is the set of commands fixed?

  • CVE-2025-54994: @akoskm/create-mcp-server-stdio is vulnerable to MCP Server Command Injection through `exec` API
  • CVE-2025-59377: mcp-kubernetes-server has an OS Command Injection vulnerability
  • CVE-2025-59834: Command Injection in adb-mcp MCP Server

Path traversal and file access · 51 (20%)

A tool takes a path and reads, writes or uploads it. Without a root it is held to, the tool reaches any file the server's account can.

Check: For every tool that takes a path: is it resolved and checked against an allowed root after symlinks, on every call?

  • CVE-2026-114125: Storybook: Missing origin validation in Storybook preview allows code injection from malicious websites
  • CVE-2026-11720: MCP Toolbox HTTP tool path parameters allow traversal to unintended endpoints
  • CVE-2026-27825: MCP Atlassian has an arbitrary file write leading to arbitrary code execution via unconstrained download_path in confluence_download_attachment

Server-side request forgery · 40 (15%)

A tool fetches a URL the caller supplies. From inside a network that reaches the cloud metadata address and internal services.

Check: For every tool that makes a request: is the destination checked after DNS resolution and on every redirect, with internal ranges refused?

  • CVE-2026-61559: @zereight/mcp-gitlab Vulnerable to Server-Side Request Forgery
  • CVE-2026-26118: Azure MCP Server has Server-Side Request Forgery issue that allows authorized attacker to elevate privileges over a network
  • CVE-2026-27826: MCP Atlassian has SSRF via unvalidated X-Atlassian-Jira-Url / X-Atlassian-Confluence-Url headers

Missing or weak authentication · 38 (15%)

An HTTP transport that answers anyone who can reach the port, or a default that binds to every interface with no credential.

Check: Does every transport other than stdio require a credential by default, and what does the server bind to out of the box?

  • CVE-2025-49596: MCP Inspector proxy server lacks authentication between the Inspector client and proxy
  • CVE-2026-11717: googleapis/mcp-toolbox: authentication bypass vulnerability in the generic opaque token validation path (validateOpaqueToken)
  • CVE-2026-48039: Meta Ads MCP: Unauthenticated HTTP MCP Tool Execution Leaks Operator Meta Access Token

Missing or wrong authorization · 35 (13%)

The server acts with one identity for every caller, or a read-only mode that some tool or procedure walks around.

Check: Whose permissions does each tool call run with, and is the caller's own access checked on each call, not once per session?

  • CVE-2026-114125: Storybook: Missing origin validation in Storybook preview allows code injection from malicious websites
  • CVE-2026-54052: n8n-MCP: Cross-tenant access to workflow version backups in multi-tenant HTTP deployments
  • CVE-2026-77244: [mcp-atlassian] Authentication bypass in HTTP transport: AtlassianOpaqueTokenVerifier accepts any non-empty token

Information exposure · 24 (9%)

Tokens, request bodies or internal errors written to logs or returned to the caller.

Check: What does the server log and return on an error, and can a credential it holds end up in either?

  • CVE-2026-48039: Meta Ads MCP: Unauthenticated HTTP MCP Tool Execution Leaks Operator Meta Access Token
  • CVE-2026-25650: MCP-Salesforce's arbitrary attribute access leads to disclosure of Salesforce auth token
  • CVE-2026-62264: @sylphx/pdf-reader-mcp: HTTP transport accepts unauthenticated requests despite a configured MCP_API_KEY

Origin checks and cross-site flaws · 21 (8%)

A local server that trusts any web page able to reach localhost: no Origin or Host check, so a page in the user's browser can call its tools.

Check: Does a local HTTP or SSE server check Origin and Host, and does any consent or admin page escape what it prints?

  • CVE-2026-114125: Storybook: Missing origin validation in Storybook preview allows code injection from malicious websites
  • CVE-2026-11624: MCP Toolbox for Databases has an Origin Validation Error
  • CVE-2026-59971: MySQL MCP Server: Missing Origin/Host Validation in SSE Transport Enables Unauthenticated SQL Execution (DNS Rebinding / Direct Exposure)

Denial of service · 14 (5%)

One request that exhausts memory or CPU: an unbounded body, a recursive structure, a slow regular expression.

Check: Are request size, depth and time bounded before parsing, and does one caller's request affect the others?

  • CVE-2025-53365: MCP Python SDK has Unhandled Exception in Streamable HTTP Transport, Leading to Denial of Service
  • CVE-2025-53366: MCP Python SDK vulnerability in the FastMCP Server causes validation error, leading to DoS
  • CVE-2026-0621: Anthropic's MCP TypeScript SDK has a ReDoS vulnerability

SQL and query injection · 4 (2%)

A tool builds a database query from its arguments, or a read-only promise enforced by inspecting the query text.

Check: Are queries parameterized, and is read-only enforced by the database account, not by parsing the statement?

  • CVE-2025-66336: Apache Doris MCP Server is vulnerable to SQL Injection via metadata query path
  • CVE-2026-33980: Azure Data Explorer MCP Server: KQL Injection in multiple tools allows MCP client to execute arbitrary Kusto queries
  • CVE-2025-66335: Apache Doris MCP Server vulnerable to SQL Injection via improper query context neutralization

When they were published

Each vulnerability is dated by its earliest public advisory. Before May 29, 2025 the record holds none for an MCP-named package.

MCP-named packagesSDKs and frameworks
One column per quarter, by the date each vulnerability was first published. * The last quarter is not over: it holds the records published so far.
QuarterMCP-named packagesSDKs and frameworks
2025 Q257
2025 Q32321
2025 Q41121
2026 Q12727
2026 Q28139
2026 Q39412
2026 Q4 (to date)2619

Where they are

  • The 267 vulnerabilities are spread over 145 MCP-named packages, and 100 of those packages (69%) have exactly one.
  • The ten packages with the most account for 89 of the 267 (33%).
  • By registry: 143 on npm, 109 on PyPI, 17 on Go.
  • For 44 (16%), the record names no version that fixes the flaw.
PackageRegistryVulnerabilities
mcp-atlassianPyPI30
n8n-mcpnpm11
mcpPyPI10
@modelcontextprotocol/sdknpm9
@aborruso/ckan-mcp-servernpm7
github.com/googleapis/mcp-toolboxGo6
mcp-server-kubernetesnpm6
mcp-searxngnpm5
@dynatrace-oss/dynatrace-mcp-servernpm4
@modelcontextprotocol/servernpm4

A long list is not a verdict on a package: projects that are widely used and run a disclosure process collect more advisories than projects nobody examines.

A flaw in a real package is one risk. A package that is itself the attack is another: the same record holds 170 reports of malicious code published under an MCP-named package. They are counted on AI agent stack vulnerabilities, not here.

Method

  • Vulnerabilities. The MCP-named subset of AI agent stack vulnerabilities: public OSV records for npm, PyPI and Go, copied on November 13, 2026, with records that share an identifier counted once.
  • Weakness type. The CWE identifiers each advisory carries in OSV's record. GitHub-reviewed advisories have them; PyPA and Go records mostly do not. 7 MCP vulnerabilities have none and are not classified.
  • Classes. A fixed table maps CWE identifiers to the ten types above; an identifier on no row is “other”. The table:
Flaw typeCWE identifiers
Command and code injection74, 77, 78, 88, 94, 95, 96, 917, 1336
Path traversal and file access22, 23, 35, 36, 59, 73, 434, 552
Server-side request forgery918
Missing or weak authentication287, 288, 290, 294, 306, 307, 345, 347, 384, 522, 613, 798, 1188
Missing or wrong authorization266, 269, 284, 285, 441, 639, 732, 749, 862, 863
Information exposure200, 201, 209, 312, 319, 359, 497, 532, 538
Origin checks and cross-site flaws79, 346, 350, 352, 601, 942, 1385
Denial of service248, 400, 405, 409, 674, 754, 770, 789, 835, 1333
Unsafe deserialization502
SQL and query injection89, 943

Who classified. The advisories, then code. No person read the individual records and no model assigned a type. Each row of the CSV carries its CWE identifiers and classes, so any figure here can be recounted.

Limits

  • These are flaws someone reported in a package manager's ecosystem. Weaknesses specific to how a model uses tools, such as a poisoned tool description, are rarely filed as package advisories and are underrepresented here.
  • A CWE identifier is the advisory author's choice. Two authors can label the same flaw differently.
  • The name rule misses MCP servers whose package name does not say MCP.
  • A class with many advisories is a class people look for and report, not proof that it is the most common flaw in servers nobody examined.

Cite this page

Drel, “MCP server vulnerabilities by flaw type,” records as of November 13, 2026, https://drel.ai/mcp-server-vulnerabilities. The figures change when the advisory copy is refreshed; the address does not. Quote the date with the number.

Advisory titles and records belong to their sources and are used under their licenses: CC BY 4.0 for the GitHub, PyPA and Go advisory databases.

Frequently asked questions

What are the most common MCP server vulnerabilities?
By the weakness types their own advisories assign, as of November 13, 2026: command and code injection (23%), path traversal and file access (20%), server-side request forgery (15%). The shares are of the 260 known MCP vulnerabilities that carry a weakness type, and one vulnerability can be in more than one.
Are MCP vulnerabilities a new kind of flaw?
In the public record, 57% of the classified MCP vulnerabilities are in three long-known classes, and 12% carry a weakness type outside the ten classes listed here. Flaws specific to how a model uses tools are rarely filed as package advisories, so this record underrepresents them.
Did Drel find or verify these vulnerabilities?
No. Every record comes from a public advisory database through OSV. Drel grouped them by the weakness identifiers the advisories carry, in code, and did not test or review the individual vulnerabilities.

Review the MCP server before the agent is given it

An AI security review covers what each tool can reach, whose authority it acts with, the attack paths through it, the controls in place and the decision your AI Committee signs.

A note on scope: Drel reviews assessed systems against documented architecture, configuration and intent. It does not ingest live telemetry from production environments. Dispositions reflect the assessed system at the time of review and the re-assessment triggers that govern when the disposition must be revisited.