fix(everything): block SSRF to internal/metadata IPs in gzip-file-as-resource

[olaservo]

Description

Adds SSRF protection to the everything server's gzip-file-as-resource tool. The tool fetched a caller-supplied http/https URL and returned the (compressed) response as a resource. Its only host restriction was an optional domain allowlist (GZIP_ALLOWED_DOMAINS) that is empty by default and treated as "all domains allowed", with no IP-range filtering and automatic redirect following. Because the URL is model-produced and prompt-injection-steerable, the server could be driven to fetch loopback, private, link-local, and cloud-metadata endpoints (e.g. 169.254.169.254) and return their contents to the caller.

Server Details

• Server: everything (TypeScript, @modelcontextprotocol/server-everything)
• Changes to: tools (gzip-file-as-resource)

Motivation and Context

validateDataURI allowed http/https/data and enforced only the domain allowlist (skipped entirely when empty), and fetchSafely called fetch(url) with default redirect following and no loopback/RFC1918/link-local/metadata checks.

What this PR does:

• Adds assertPublicHost() plus IPv4/IPv6 classifiers that resolve the destination host and refuse non-public addresses (loopback, private/RFC1918, link-local/metadata, ULA, multicast, reserved, unspecified), covering IPv4, IPv6, and IPv4-mapped IPv6 (including the hex form the WHATWG URL parser normalizes to).
• Replaces automatic redirect following with fetchWithGuardedRedirects (redirect: "manual"), re-validating the host on every hop and refusing redirects to non-http(s) schemes.
• The IP guard applies regardless of GZIP_ALLOWED_DOMAINS. The domain-allowlist semantics are intentionally left unchanged (empty still means "all domains"), so the tool's default demo URL keeps working; the allowlist and the SSRF guard are independent controls.

How Has This Been Tested?

• New unit tests assert the tool refuses non-public hosts (loopback, 169.254.169.254, RFC1918 ranges, 0.0.0.0, IPv6 loopback, IPv4-mapped IPv6) via IP literals (no network required).
• npm run build (tsc) passes; full vitest suite passes (115 tests, including the 8 new cases).
• Existing data: URI and public-URL behavior is unchanged and still covered by tests.

Breaking Changes

No client configuration changes required. Behavior only changes for requests that targeted non-public IPs, which are now refused (the intended fix). data: URIs and public URLs are unaffected.

Types of changes

• Bug fix (non-breaking change which fixes an issue)
• New feature (non-breaking change which adds functionality)
• Breaking change (fix or feature that would cause existing functionality to change)
• Documentation update

Checklist

• I have read the MCP Protocol Documentation
• My changes follows MCP security best practices
• I have updated the server's README accordingly
• I have tested this with an LLM client
• My code follows the repository's style guidelines
• New and existing tests pass locally
• I have added appropriate error handling
• I have documented all environment variables and configuration options

Additional context

Docs updated in docs/instructions.md and docs/structure.md to note that internal/metadata IPs are always blocked (and re-validated across redirects) independent of the allowlist. As with any resolve-then-connect guard, a narrow DNS-rebinding window remains; pinning to the validated IP could be added later.

🤖 Generated with Claude Code