Full Disclosure mailing list archives

-----BEGIN SECURITY ADVISORY-----

Title: Server-Side Request Forgery (SSRF) in Anthropic mcp-server-fetch and Microsoft playwright-mcp
Author: Syed Anas Mohiuddin <anasmohiuddinsyed () gmail com>
Date: May 25, 2026
CVSS: 7.5 (HIGH) — AV:N/AC:L/PR:N/UI:R/S:U/C:H/I:N/A:N
References: Already public via GitHub issues (see below)

== AFFECTED PRODUCTS ==

1. Anthropic mcp-server-fetch (modelcontextprotocol/servers)
   All versions as of May 2026
   GitHub: https://github.com/modelcontextprotocol/servers
   Public issues: #4116, #4143, #4205

2. Microsoft playwright-mcp
   All versions as of May 2026
   GitHub: https://github.com/microsoft/playwright-mcp
   Public issue: #1626

== VULNERABILITY DESCRIPTION ==

Both MCP servers accept arbitrary URLs passed by the AI agent/client without
any allowlist enforcement, IP range blocking, or internal network filtering.
This enables Server-Side Request Forgery (SSRF) attacks via prompt injection:

Attack chain:
  1. Attacker embeds malicious instruction in a webpage
  2. AI agent fetches the page via mcp-server-fetch or playwright-mcp
  3. Embedded instruction redirects the agent to fetch the cloud metadata endpoint
  4. Agent calls fetch_url("http://169.254.169.254/latest/meta-data/iam/security-credentials/";)
  5. IMDSv1 returns IAM credentials without authentication
  6. Agent includes credentials in its next response
  7. Credentials exfiltrated

Additional finding in mcp-server-fetch:
The get_prompt handler calls fetch_url() directly without invoking
check_may_autonomously_fetch_url(), bypassing the robots.txt autonomy guard
through a structurally distinct code path (logic bypass).

== DISCOVERY ==

Found using mcp-safeguard, an open-source automated security scanner for MCP servers.
pip install mcp-safeguard
https://pypi.org/project/mcp-safeguard/

Scanning 54 production MCP servers: 27.8% had HIGH/CRITICAL findings.
8/54 (14.8%) confirmed SSRF. 7/54 credential exposure.

== DISCLOSURE TIMELINE ==

May 2026: Findings discovered via mcp-safeguard
May 2026: Reported to Anthropic Security (security () anthropic com)
May 2026: Reported to Microsoft MSRC (secure () microsoft com)
May 2026: Issues already publicly visible on GitHub (see References above)
May 2026: Public advisory posted to Full Disclosure

== MITIGATIONS ==

For MCP server operators:
- Enforce URL allowlists (only fetch from approved domains)
- Block RFC1918 and link-local ranges at the application layer
- Use IMDSv2 (requires session token; not fetchable via simple HTTP)
- Pin resolved IPs before making TCP connections (prevents DNS rebinding)
- Validate redirect destinations before following

For AI agent deployments:
- Review all MCP servers in your stack using mcp-safeguard
- Apply network-level SSRF mitigations (cloud security groups, VPC policies)
- Disable IMDSv1 on all EC2 instances

== REFERENCES ==

Public GitHub issues (already disclosed):
- https://github.com/modelcontextprotocol/servers/issues/4116
- https://github.com/modelcontextprotocol/servers/issues/4143
- https://github.com/modelcontextprotocol/servers/issues/4205
- https://github.com/microsoft/playwright-mcp/issues/1626

Protocol Pivoting preprint (cross-protocol attack escalation):
https://zenodo.org/records/20371152

mcp-safeguard (detection tool):
https://pypi.org/project/mcp-safeguard/

-----END SECURITY ADVISORY-----

Syed Anas Mohiuddin
AI Security Researcher
anasmohiuddinsyed () gmail com
_______________________________________________
Sent through the Full Disclosure mailing list
https://nmap.org/mailman/listinfo/fulldisclosure
Web Archives & RSS: https://seclists.org/fulldisclosure/

Current thread:

• SSRF in Anthropic mcp-server-fetch and Microsoft playwright-mcp — publicly disclosed via GitHub issues outreach (May 25)