Render, MCP, command, and HTTP-boundary batch¶
Source: GitHub Security Advisories updated 2026-05-13.
This batch is durable because small helper features — content negotiation, cookie helpers, metadata tag names, and MCP HTTP endpoints — can quietly cross into browser script execution, HTTP response splitting, shell argument injection, or privileged tool invocation.
Advisories covered¶
- Fiber AutoFormat XSS — GHSA-qjv7-627w-8qjv: content negotiation/AutoFormat could emit scriptable output. Fixed in
fiber/v2 2.52.13andfiber/v3 3.2.0. - Microdot response splitting — GHSA-7wc8-wvc4-m498:
Response.set_cookie()accepted values that could split HTTP responses. Fixed in2.6.1. - exiftool-vendored argument injection — GHSA-cw26-7653-2rp5: newline characters in tag names reached ExifTool arguments. Fixed in
35.19.0. - Network-AI unauthenticated MCP HTTP endpoint — GHSA-fj4g-2p96-q6m3: missing auth allowed unauthenticated privileged tool calls. Fixed in
5.1.3.
Operator triage¶
- Patch public Fiber and Microdot services that reflect negotiated output or set attacker-influenced cookies.
- Search logs for response-splitting probes:
%0d%0a, raw CR/LF, duplicate headers, and injectedSet-Cookie/Locationheaders. - Patch
exiftool-vendored; review upload pipelines for attacker-controlled tag names, newline-delimited arguments, and ExifTool invocation logs. - Treat unauthenticated MCP HTTP endpoints as exposed admin planes. Disable them or bind to loopback until auth, origin, and network policy are verified.
Durable controls¶
- Content negotiation must preserve output context: reflected data needs the encoder for the selected media type, not a generic serializer.
- Cookie/header helper APIs should reject CTLs, CR/LF, and invalid delimiters before constructing response headers.
- CLI wrappers must pass structured arguments with strict allowlists; metadata keys are not safe flag or argument names.
- MCP servers need explicit authentication, origin policy, audit logging, and least-privilege tool scopes even on “local” HTTP transports.