Don't resolve or dial the origin ourselves when a proxy is set

The connect-fallback probe resolved the origin locally and, on a proxy-connect
failure, fell back to dialing the origin direct — bypassing the proxy, leaking
the origin's DNS query, and sending it the request (with the full absolute URI)
over a direct connection. Under a proxy the socket dials the proxy, not the
origin, so an origin-address list is meaningless here; gate the probe on
!proxy.active. A dead proxy now fails cleanly instead of falling open to the
origin.

Test 55 drives a crawl through a dead proxy at a two-address origin decoy and
asserts the decoy is never contacted; it fails on the pre-fix engine.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Signed-off-by: Xavier Roche <roche@httrack.com>

.github/workflows/codeql.yml

@@ -46,6 +46,12 @@ jobs:
           # fopen's umask-controlled 0666 is intended for mirror/cache/log
           # output; the one credential file (cookies.txt) is kept 0600 on Unix.
           config: |
+            paths-ignore:
+              # Demo callback samples, not part of libhttrack.
+              - libtest
+              # ProxyTrack: a separate legacy binary with no auth surface; its
+              # recv/cache-parse code trips taint queries by design.
+              - src/proxy
             query-filters:
               - exclude:
                   id: cpp/world-writable-file-creation

src/htsback.c

@@ -2696,9 +2696,12 @@ void back_wait(struct_back * sback, httrackp * opt, cache_back * cache,
         int dispo = 0;
 
         // probe the resolved address list once per fresh connect (cache hit:
-        // the host was resolved when this connect was opened)
+        // the host was resolved when this connect was opened). Not under a
+        // proxy: the socket dials the proxy, so resolving the origin here leaks
+        // its DNS and lets a proxy-connect failure fall back to dialing it
+        // direct.
         if (cf->addr_count < 0 && back[i].r.soc != INVALID_SOCKET &&
-            !back[i].r.is_file) {
+            !back[i].r.is_file && !back[i].r.req.proxy.active) {
           SOCaddr scratch[HTS_MAXADDRNUM];
 
           cf->addr_count = hts_dns_resolve_all(opt, back[i].url_adr, scratch,

tests/56_local-proxy-noleak.test

@@ -0,0 +1,93 @@
+#!/bin/bash
+#
+# A crawl through a proxy must never resolve or dial the origin itself: the proxy
+# does that. A dead proxy + a multi-address origin used to fall back to dialing
+# the origin direct (bypassing the proxy, leaking its DNS and IP). The decoy
+# origin here must therefore receive nothing.
+
+set -euo pipefail
+
+: "${top_srcdir:=..}"
+
+# shellcheck source=tests/testlib.sh
+. "$top_srcdir/tests/testlib.sh"
+
+if test "${V6_SUPPORT:-}" == "no"; then
+    echo "no IPv6 support (resolver override compiled out), skipping"
+    exit 77
+fi
+python=$(find_python) || {
+    echo "python3 missing, skipping"
+    exit 77
+}
+
+server=$(nativepath "$top_srcdir/tests/local-server.py")
+root=$(nativepath "$top_srcdir/tests/server-root")
+tmpdir=$(mktemp -d)
+serverpid=
+
+cleanup() {
+    stop_server "$serverpid"
+    rm -rf "$tmpdir"
+    return 0
+}
+trap cleanup EXIT
+
+# decoy origin: it must stay silent. LOCAL_SERVER_VERBOSE logs any request it gets.
+LOCAL_SERVER_VERBOSE=1 "$python" "$server" --root "$root" --bind 127.0.0.1 \
+    >"$tmpdir/srv.out" 2>"$tmpdir/srv.err" &
+serverpid=$!
+port=
+for _ in $(seq 1 50); do
+    line=$(head -n1 "$tmpdir/srv.out" 2>/dev/null || true)
+    if test "${line%% *}" == "PORT"; then
+        port="${line#PORT }"
+        break
+    fi
+    kill -0 "$serverpid" 2>/dev/null || {
+        echo "server exited early: $(cat "$tmpdir/srv.err")"
+        exit 1
+    }
+    sleep 0.1
+done
+test -n "$port" || {
+    echo "could not discover server port"
+    exit 1
+}
+
+# a proxy port nothing listens on: grab a free one and let it close (refuses).
+deadproxy=$("$python" -c \
+    'import socket; s=socket.socket(); s.bind(("127.0.0.1",0)); print(s.getsockname()[1]); s.close()')
+
+# origin resolves to two addresses so a fallback would have a second to dial;
+# 127.0.0.1 (the decoy) is the one the old bypass reached.
+out="$tmpdir/crawl"
+HTTRACK_DEBUG_RESOLVE="decoyhost:127.0.0.2,127.0.0.1" \
+    httrack "http://decoyhost:$port/simple/basic.html" -O "$out" \
+    -P "127.0.0.1:$deadproxy" -c1 --robots=0 --timeout=5 --retries=0 --quiet -Z \
+    >"$tmpdir/log" 2>&1
+
+log="$out/hts-log.txt"
+
+# the origin must have seen no connection (no leak/bypass)
+if grep -qE '"(GET|POST|HEAD|CONNECT) ' "$tmpdir/srv.err"; then
+    echo "FAIL: origin was contacted directly, bypassing the proxy"
+    cat "$tmpdir/srv.err"
+    exit 1
+fi
+
+# the crawl must fail at the proxy (proves it really tried the proxy, not a no-op)
+grep -q 'Connect Error' "$log" || {
+    echo "FAIL: expected a proxy connect error"
+    cat "$log"
+    exit 1
+}
+
+# and it must not have tried an origin-address fallback under the proxy
+if grep -q "trying next address" "$log"; then
+    echo "FAIL: fell back to an origin address under a proxy"
+    cat "$log"
+    exit 1
+fi
+
+echo "OK: a dead proxy fails cleanly without dialing the origin"

tests/Makefile.am

@@ -132,6 +132,7 @@ TESTS = \
 	52_local-socks5.test \
 	53_local-proxytrack-cache-corrupt.test \
 	54_local-update-truncate-purge.test \
-	55_local-chunked.test
+	55_local-chunked.test \
+	56_local-proxy-noleak.test
 
 CLEANFILES = check-network_sh.cache