Hackney has SSRF allowlist bypass in hackney_url:normalize/2 via percent-encoded host

Summary

hackney_url:normalize/2 URL-decodes the host component of a parsed URL, but the caller's SSRF allowlist runs before normalization using OTP's uri_string:parse/1 and inet:parse_address/1, neither of which decodes percent-escapes in hostnames. A URL like http://%31%32%37%2E%30%2E%30%2E%31/ presents an encoded, non-IP-looking host to the validator, which passes the allowlist check; hackney's normalizer then decodes it to 127.0.0.1 and connects to loopback. Because hackney:request/5 always calls normalize/2 with no opt-out, every request path that accepts a binary or list URL is affected. This is a parser-differential SSRF in the same class as CVE-2025-1211, but in a different function.

Details

In src/hackney_url.erl (lines 161–186), normalize/2 checks whether the parsed host is already a dotted-quad or IPv6 literal via inet_parse:address/1. Percent-encoded forms like %31%32%37%2E%30%2E%30%2E%31 fail that check and fall into the catch-all branch, where urldecode/1 decodes the host before passing it to IDNA conversion:

Host1 = binary_to_list(
           urldecode(unicode:characters_to_binary(Host0))
         ),

The decoded host ("127.0.0.1") replaces the original in the returned #hackney_url{} record. hackney:request/5 at src/hackney.erl:463 always calls normalize/2, so the decoded host is what do_dispatch/1 and add_host_header/2 ultimately use. The on-wire Host: header and the TCP connect target both reflect the decoded value.

The same payload pattern reaches the AWS/GCP/Azure IMDS (169.254.169.254), RFC1918 ranges, and any localhost admin endpoint. The 1.21.0 patch for CVE-2025-1211 fixed a separate differential in parse_url/1 and did not touch normalize/2.

PoC

1. Validate the URL with the canonical Erlang SSRF allowlist: uri_string:parse/1 returns host , returns , so the allowlist accepts it.