What happened
NVD published CVE-2026-19039, CVE-2026-19044, and CVE-2026-19038 (CVSS 5.3, Medium) on August 6, 2026, covering command injection and path traversal flaws in three niche, single-author MCP server integrations that bridge AI agents to SSH, Godot, and browser screenshot tooling.
Why it matters
These represent the long tail of the rapidly-growing MCP ecosystem: low-maintenance, single-author tool bridges that grant agents direct system access (SSH execution, local file writes) with inadequate input sanitization — precise, low-blast-radius but indicative of a systemic pattern across the MCP tool supply chain.
Attack vector
ssh_exec in ssh-mcp-server allows command injection via host/username arguments; godot-mcp's executeOperation allows command injection via projectPath (local attack only); la-forge-mcp's screenshotElement allows path traversal via output_name.
Affected systems
Kino-Kafkaesque ssh-mcp-server (pre-fix commit), LeeSinLiang godot-mcp 0.1.0, MonomythDevelopment la-forge-mcp 1.0.0
Mitigation
Upgrade to patched releases of each respective MCP server; avoid exposing local MCP tool bridges to untrusted input.