Vulnerabilities in agent tooling
Public, sourced advisories for MCP servers and agent tools, in the OSV format with VERFY-YYYY-NNNN IDs and CVE aliases. Feed: /api/v1/advisories.
MCP tool definitions change between releases, often in patch releases
We installed the newest releases of 46 popular MCP servers on throwaway machines and compared the tools each release served. Of the 34 that started, 28 added, removed or changed a tool definition, and 23 of 70 patch releases changed one: a version range like ^1.2.3 does not freeze what your agent reads. We found no malicious change. The same run showed 163 false alarms in Verfy's own scanner, which 0.5.2 fixes.
Every server, release by release · Method, numbers and limits · Code to re-run it
VERFY-2025-0001CVE-2025-6514source: public recordmcp-remote: OS command injection from a malicious MCP server's OAuth authorization_endpoint
npm · mcp-remote · added
VERFY-2025-0002CVE-2025-49596source: public recordMCP Inspector: unauthenticated proxy endpoints allow remote code execution
npm · @modelcontextprotocol/inspector · added
VERFY-2025-0003CVE-2025-53109source: public recordMCP filesystem server: symlinks inside an allowed directory escape the allowed-directory boundary
npm · @modelcontextprotocol/server-filesystem · added
VERFY-2025-0004CVE-2025-53110source: public recordMCP filesystem server: path prefix collision bypasses the allowed-directory check
npm · @modelcontextprotocol/server-filesystem · added