Severity by source
CVSS:4.0/AV:N/AC:H/AT:N/PR:N/UI:A/VC:H/VI:H/VA:N/SC:N/SI:N/SA:N/E:X/CR:X/IR:X/AR:X/MAV:X/MAC:X/MAT:X/MPR:X/MUI:X/MVC:X/MVI:X/MVA:X/MSC:X/MSI:X/MSA:X/S:X/AU:X/R:X/V:X/RE:X/U:X
Network attacker needs no privileges (PR:N) but must socially engineer a victim into completing an OAuth flow (UI:R, AC:H); stolen token grants access to a separate server, so S:C with high C/I impact and no availability effect.
Primary rating from Vendor (https://github.com/PrefectHQ/fastmcp).
CVSS VectorVendor: https://github.com/PrefectHQ/fastmcp
CVSS:4.0/AV:N/AC:H/AT:N/PR:N/UI:A/VC:H/VI:H/VA:N/SC:N/SI:N/SA:N/E:X/CR:X/IR:X/AR:X/MAV:X/MAC:X/MAT:X/MPR:X/MUI:X/MVC:X/MVI:X/MVA:X/MSC:X/MSI:X/MSA:X/S:X/AU:X/R:X/V:X/RE:X/U:X
Lifecycle Timeline
12Blast Radius
ecosystem impact- 21 pypi packages depend on fastmcp (20 direct, 1 indirect)
Ecosystem-wide dependent count for version 2.14.2.
DescriptionCVE.org
While testing the OAuth Proxy implementation, it was noticed that the server does not properly respect the resource parameter submitted by the client in the authorization and token request. Instead of issuing the token explicitly for this MCP server, the token is issued for the base_url passed to the OAuthProxy during initialization.
Affected File: *https://github.com/jlowin/fastmcp/blob/main/src/fastmcp/server/auth/oauth_proxy.py#L828*
Affected Code:
self._jwt_issuer: JWTIssuer = JWTIssuer(
issuer=str(self.base_url),
audience=f"{str(self.base_url).rstrip('/')}/mcp",
signing_key=jwt_signing_key,
)Since the issued access and refresh tokens do not include information about the resource the token was issued for, it is impossible for the MCP server to properly verify whether the token was issued for it, hence violating the requirement of doing so demanded by the specification. Being able to verify whether the token was issued for the target MCP server enforces the protection offered by the, per MCP specification mandatory, Protected Resource Metadata OAuth extension.
Therefore, this misconfiguration exposes all MCP server setups using the FastMCP OAuth Proxy to an attack where an adversary creates a malicious MCP server that advertises the benign OAuth Proxy authorization server as its own authorization server. Once a victim completes an OAuth flow with this malicious MCP server, authenticating against the AS, the adversary can extract the token received at the malicious MCP server and use it to access other MCP servers (the benign ones) that also use the same AS, including the tools and resources they expose.
Steps to reproduce:
- Extract the provided PoC environment.
- Enter the *client_id* and *client_secret* of a GitHub App you control into the
mcp-server-proxy.pyscript. - Start the benign MCP server using an OAuth Proxy (in this case the *GitHubProvider*):
python3 mcp-server-proxy.py. - Start the malicious AS:
python3 mal_auth_server.py. - Start the malicious MCP server:
python3 attacker_server.py. - Connect the client to the malicious MCP server:
python3 client.py. - Complete the OAuth flow.
- Observe in the logs of the malicious MCP server that the request to the benign MCP server with the stolen token returned a 200 status code.
Impact
This vulnerability allows an adversary to steal a victim’s authentication material for a benign MCP server using the FastMCP OAuth Proxy. The severity of this issue was decreased to _Medium_ due to the consent screen showing the name of the MCP server the OAuth Proxy was intended for. However, a victim might not see it or get otherwise convinced by the attacker to ignore it, and overall this does not act as a proper mitigation for this issue.
Mitigation
To mitigate this vulnerability, it is recommended to issue tokens specifically for the MCP server submitted in the authorization URL’s resource GET parameter. In this way, the receiving MCP server will be able to properly verify that the token was indeed issued for it, allowing it to reject tokens stolen by an attack like the one demonstrated above.
AnalysisAI
Token audience confusion in the FastMCP OAuth Proxy (fastmcp < 2.14.2) lets a malicious MCP server harvest a victim's access/refresh tokens and replay them against benign MCP servers that share the same authorization server. The proxy ignores the client-supplied resource parameter and instead binds every token to the proxy's own base_url, so receiving servers cannot enforce audience validation. A working proof-of-concept environment is published, though EPSS is negligible (0.01%) and the issue is not in CISA KEV; publicly available exploit code exists.
Technical ContextAI
FastMCP is a Python framework (pkg:pip/fastmcp) for building Model Context Protocol servers; its OAuthProxy component fronts an upstream OAuth authorization server (e.g. GitHubProvider) and issues JWTs to MCP clients. In src/fastmcp/server/auth/oauth_proxy.py the JWTIssuer is initialized with issuer=base_url and audience=base_url/mcp, hard-coding the audience to the proxy's configured base_url rather than the per-request resource value mandated by the MCP authorization spec (RFC 8707 Resource Indicators / token audience binding). This is a CWE-863 (Incorrect Authorization) flaw: because the token never records which resource it was minted for, any MCP server validating it cannot distinguish a token issued for itself from one issued for a different server using the same AS, defeating the Protected Resource Metadata protection.
RemediationAI
Vendor-released patch: upgrade fastmcp to 2.14.2 or later, which is the fixed release per the GitHub advisory GHSA-5h2m-4q8j-pqpj and EUVD-2025-208759; the fix issues tokens bound to the resource value from the authorization URL so receiving servers can validate the audience. If you cannot upgrade immediately, the practical compensating control is to avoid sharing a single FastMCP OAuth Proxy authorization server across multiple independently-trusted MCP servers, so a token stolen at one server is not accepted by another (trade-off: loses SSO convenience across your MCP fleet). Operationally, instruct users to scrutinize the OAuth consent screen, which shows the name of the intended MCP server, and to abort flows initiated from untrusted servers - note this is a weak control, as the reporter explicitly states the consent screen does not constitute a proper mitigation. Advisory and details: https://github.com/PrefectHQ/fastmcp/security/advisories/GHSA-5h2m-4q8j-pqpj.
Wazuh SIEM platform versions 4.4.0 through 4.9.0 contain an unsafe deserialization vulnerability in the DistributedAPI t
BentoML version 1.4.2 and earlier contains an unauthenticated remote code execution vulnerability through insecure deser
pgAdmin 4 contains critical remote code execution vulnerabilities in the Query Tool download and Cloud Deployment endpoi
The renderLocalView function in render/views.py in graphite-web in Graphite 0.9.5 through 0.9.10 uses the pickle Python
BentoML is a Python library for building online serving systems optimized for AI apps and model inference. Rated critica
OpenSSL before 0.9.8za, 1.0.0 before 1.0.0m, and 1.0.1 before 1.0.1h does not properly restrict processing of ChangeCiph
pyLoad download manager version prior to 0.5.0b3.dev77 exposes the Flask SECRET_KEY through an unauthenticated endpoint.
In Mercurial before 4.1.3, "hg serve --stdio" allows remote authenticated users to launch the Python debugger, and conse
Unauthenticated remote code execution in Marimo ≤0.20.4 allows attackers to execute arbitrary system commands via the `/
pyLoad is the free and open-source Download Manager written in pure Python. Rated medium severity (CVSS 5.3), this vulne
Langflow (a visual LLM pipeline builder) contains a critical unauthenticated code execution vulnerability (CVE-2026-3301
Cross-user flow execution in Langflow (< 1.9.1) lets any authenticated API-key holder run another user's flow by passing
Same weakness CWE-863 – Incorrect Authorization
View allSame technique Authentication Bypass
View allVendor StatusVendor
Share
External POC / Exploit Code
Leaving vuln.today
EUVD-2025-208759
GHSA-5h2m-4q8j-pqpj