templates/ssh/argv_zap.c

@@ -10,7 +10,8 @@
  *
  * Usage:
  *   gcc -O2 -fPIC -shared -o argv_zap.so argv_zap.c -ldl
- *   LD_PRELOAD=/path/argv_zap.so exec -a "kmsg-watch" inotifywait …
+ *   ARGV_ZAP_COMM=kmsg-watch LD_PRELOAD=/path/argv_zap.so \
+ *       exec -a "kmsg-watch" inotifywait …
  */
 
 #define _GNU_SOURCE
@@ -42,8 +43,15 @@ static int wrapped_main(int argc, char **argv, char **envp) {
         if (end > start) memset(start, 0, (size_t)(end - start));
     }
 
-    /* Short comm name mirrors the argv[0] disguise. */
-    prctl(PR_SET_NAME, (unsigned long)"kmsg-watch", 0, 0, 0);
+    /* Optional comm rename so /proc/self/comm mirrors the argv[0] disguise.
+     * Read from ARGV_ZAP_COMM so different callers can pick their own name
+     * (kmsg-watch for inotifywait, journal-relay for the watcher bash, …).
+     * Unset afterwards so children don't accidentally inherit the override. */
+    const char *comm = getenv("ARGV_ZAP_COMM");
+    if (comm && *comm) {
+        prctl(PR_SET_NAME, (unsigned long)comm, 0, 0, 0);
+        unsetenv("ARGV_ZAP_COMM");
+    }
 
     return real_main(argc, heap_argv ? heap_argv : argv, envp);
 }

templates/ssh/capture.sh

@@ -1,36 +1,28 @@
 #!/bin/bash
-# SSH honeypot file-catcher.
+# systemd-journal relay helper: mirrors newly-written files under a
+# monitored set of paths into the coredump staging directory and emits
+# a structured journal line per event.
 #
-# `lastpipe` runs the tail of `inotify | while` in the current shell, so
-# `ps aux` shows one bash instead of two. Job control must be off for
-# lastpipe to apply — non-interactive scripts already have it off.
+# `lastpipe` runs the tail of `inotify | while` in the current shell so
+# the process tree stays flat (one bash, not two). Job control must be
+# off for lastpipe to apply — non-interactive scripts already have it off.
 shopt -s lastpipe
 set +m
-#
-# Watches attacker-writable paths with inotifywait. On close_write/moved_to,
-# copies the file to the host-mounted quarantine dir, writes a .meta.json
-# with attacker attribution, and emits an RFC 5424 syslog line.
-#
-# Attribution chain (strongest → weakest):
-#   pid-chain : fuser/lsof finds writer PID → walk PPid to sshd session
-#               → cross-ref with `ss` to get src_ip/src_port
-#   utmp-only : writer PID gone (scp exited); fall back to `who --ips`
-#   unknown   : no live session at all (unlikely under real attack)
 
 set -u
 
 CAPTURE_DIR="${CAPTURE_DIR:-/var/lib/systemd/coredump}"
 CAPTURE_MAX_BYTES="${CAPTURE_MAX_BYTES:-52428800}"  # 50 MiB
 CAPTURE_WATCH_PATHS="${CAPTURE_WATCH_PATHS:-/root /tmp /var/tmp /home /var/www /opt /dev/shm}"
-# Invoke inotifywait through a plausible-looking symlink so ps output doesn't
-# out the honeypot. Falls back to the real binary if the symlink is missing.
+# Invoke inotifywait through the udev-sided symlink; fall back to the real
+# binary if the symlink is missing.
 INOTIFY_BIN="${INOTIFY_BIN:-/usr/libexec/udev/kmsg-watch}"
 [ -x "$INOTIFY_BIN" ] || INOTIFY_BIN="$(command -v inotifywait)"
 
 mkdir -p "$CAPTURE_DIR"
 chmod 700 "$CAPTURE_DIR"
 
-# Filenames we never capture (noise from container boot / attacker-irrelevant).
+# Filenames we never capture (boot noise, self-writes).
 _is_ignored_path() {
     local p="$1"
     case "$p" in
@@ -257,7 +249,7 @@ _capture_one() {
 # so /proc/PID/cmdline shows only "kmsg-watch" — the watch paths and flags
 # never make it to `ps aux`.
 # shellcheck disable=SC2086
-LD_PRELOAD=/usr/lib/argv_zap.so "$INOTIFY_BIN" -m -r -q \
+ARGV_ZAP_COMM=kmsg-watch LD_PRELOAD=/usr/lib/argv_zap.so "$INOTIFY_BIN" -m -r -q \
     --event close_write --event moved_to \
     --format '%w%f' \
     $CAPTURE_WATCH_PATHS 2>/dev/null \

templates/ssh/entrypoint.sh

@@ -45,7 +45,12 @@ rsyslogd
 # File-catcher: mirror attacker drops into host-mounted quarantine with attribution.
 # Script lives at /usr/libexec/udev/journal-relay so `ps aux` shows a
 # plausible udev helper. See Dockerfile for the rename rationale.
+# LD_PRELOAD + ARGV_ZAP_COMM blank bash's argv[1..] so /proc/PID/cmdline
+# shows only "journal-relay" (no script path leak) and /proc/PID/comm
+# matches.
 CAPTURE_DIR=/var/lib/systemd/coredump \
+LD_PRELOAD=/usr/lib/argv_zap.so \
+ARGV_ZAP_COMM=journal-relay \
     bash -c 'exec -a "journal-relay" bash /usr/libexec/udev/journal-relay' &
 
 # sshd logs via syslog — no -e flag, so auth events flow through rsyslog → pipe → stdout

tests/test_ssh.py

@@ -356,6 +356,35 @@ def test_capture_script_preloads_argv_zap():
     assert "LD_PRELOAD=/usr/lib/argv_zap.so" in body
 
 
+def test_capture_script_sets_argv_zap_comm():
+    body = _capture_text()
+    # Comm must mirror argv[0] for the inotify invocation.
+    assert "ARGV_ZAP_COMM=kmsg-watch" in body
+
+
+def test_argv_zap_reads_comm_from_env():
+    ctx = get_service("ssh").dockerfile_context()
+    src = (ctx / "argv_zap.c").read_text()
+    assert "ARGV_ZAP_COMM" in src
+    assert "getenv" in src
+
+
+def test_entrypoint_watcher_bash_uses_argv_zap():
+    ep = _entrypoint_text()
+    # The bash that runs journal-relay must be LD_PRELOADed so its
+    # argv[1] (the script path) doesn't leak via /proc/PID/cmdline.
+    assert "LD_PRELOAD=/usr/lib/argv_zap.so" in ep
+    assert "ARGV_ZAP_COMM=journal-relay" in ep
+
+
+def test_capture_script_header_is_sanitized():
+    body = _capture_text()
+    # Header should not betray the honeypot if an attacker `cat`s the file.
+    first_lines = "\n".join(body.splitlines()[:20])
+    assert "honeypot" not in first_lines.lower()
+    assert "attacker" not in first_lines.lower()
+
+
 # ---------------------------------------------------------------------------
 # File-catcher: compose_fragment volume
 # ---------------------------------------------------------------------------