An open-source scanner named mcp-witness has identified six vulnerabilities across six Python MCP servers published on PyPI. These vulnerabilities fall into two categories: outbound SSRF in URL-fetching tools and inbound DNS rebinding on HTTP-transport MCP servers. The SSRF vulnerability, found in Anthropic's reference mcp-server-fetch and mcp-server-http-request, could allow an agent to retrieve cloud-metadata-service credentials. The DNS rebinding vulnerability, present in four other packages, could enable any webpage to issue tool calls against locally running MCP servers. Despite responsible disclosure and a proposed fix, some vulnerabilities remain unaddressed. AI
IMPACT Highlights critical security flaws in AI agent transport layers, potentially impacting secure tool invocation and credential management.
RANK_REASON Disclosure of multiple vulnerabilities in open-source software by an independent researcher. [lever_c_demoted from research: ic=1 ai=1.0]
- Anthropic
- Azure
- fastmcp-http
- Google Cloud Platform
- MCP
- mcp-fetch-streamablehttp-server
- mcp-server-fetch
- mcp-server-fetch-sse
- mcp-server-http-request
- mcp-streamablehttp-proxy
- mcp-witness
- Python
- Python Package Index
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →