FastGPT AI Agent Platform Hit with Four CVEs: Unauthenticated SSRF, MCP Tool URL Bypass, DNS Rebinding, and Worker Pool DoS

TheHackerWire·medium signal

Four vulnerabilities were disclosed in FastGPT, an AI agent building platform, all patched in version 4.14.17. CVE-2026-44286 allows unauthenticated SSRF via the lafModule workflow node's fetchData function bypassing isInternalAddress(). CVE-2026-44284 exposes an SSRF gap in MCP tool URL handling. CVE-2026-42344 exploits DNS rebinding between validation and HTTP request. CVE-2026-42343 enables complete DoS by exhausting the JavaScript worker pool in the code-sandbox component. The MCP tool URL bypass is particularly notable as it represents another instance of MCP as an architectural perimeter bypass vector.