claude.ai code-execution sandbox: HTTP 503 'DNS cache overflow' on egress to Vercel-hosted host

Note: this is a bug in claude.ai's code-execution (bash) sandbox, not the Claude Code CLI. Filing here because there's no public claude.ai issue tracker — please redirect if there's a better channel.

Summary

Outbound HTTPS requests from the claude.ai code-execution sandbox to a Vercel-hosted host (shelf.grafals.net, CNAME cname.vercel-dns.com. → 76.76.21.142, 66.33.60.66) return:

HTTP/2 503
content-length: 18
content-type: text/plain

DNS cache overflow

The same request from any other source (a personal Mac on residential ISP, a Hetzner VPS) returns the expected HTTP/2 307 redirect to /login with server: Vercel. So the failure is specific to the sandbox's egress path, not the destination.

Repro

From a claude.ai code-execution bash session:

curl -sS -i "https://shelf.grafals.net/"

Expected: HTTP/2 307 redirect (Vercel responding normally).
Actual: HTTP/2 503 with body DNS cache overflow.

Reproduces consistently over a span of minutes, on retry, and across multiple *.grafals.net hosts that resolve to Vercel anycast.

What I ruled out

• Not the destination. Two unrelated source IPs (Mac, VPS) get clean 307s from shelf.grafals.net at the same time the sandbox is getting 503.
• Not our reverse proxy. shelf.grafals.net is on Vercel anycast (CNAME cname.vercel-dns.com.); my own VPS reverse proxies (Caddy/nginx) aren't in the traffic path and don't contain the string \"DNS cache overflow\" anywhere in their configs.
• Not Vercel emitting the error. The 18-byte plaintext body and the literal phrase \"DNS cache overflow\" don't match Vercel's branded error responses. Vercel uses HTML or JSON error pages with x-vercel-id / x-vercel-cache headers; the 503 here has none of those (only content-length, content-type, date).
• Not Anthropic egress allowlist. No x-deny-reason header on the response. The host is reachable enough to complete TLS and exchange HTTP/2 frames.

Hypothesis

The error string and shape look like an internal sandbox-egress proxy resolver error — the egress proxy's DNS cache filling up and emitting a custom 503 instead of resolving the destination. Pointing this out because:

• It's a custom error string (not standard HTTP, not Vercel, not Cloudflare, not standard nginx/envoy).
• It surfaces only from sandbox-origin traffic.
• TLS completes (so the traffic is reaching something on Anthropic's path, not failing at the network layer).

Why this matters

It breaks any MCP-server integration that mints signed URLs and expects the model to curl raw bytes to them from the sandbox — a common pattern for moving large payloads (audio, PDFs, screenshots) without round-tripping bytes through model context. MCP traffic itself is unaffected (different relay path), so the failure mode is asymmetric and surprising.

Suggested next steps for triage

• Grep the sandbox egress proxy config / source for the literal string \"DNS cache overflow\".
• Check resolver cache size / eviction behavior under load — the phrasing suggests a bounded LRU that returns this verbatim on overflow.
• If reproducible: confirm whether it's bound to specific destination ASNs (Vercel anycast AS13335/AS396982) or general.