feat(sniffer): capture SSH client banner from TCP stream

Parse RFC 4253 §4.2 identification strings from the first attacker→decky
data segment on TCP/22; emit ssh_client_banner syslog events and bus
fan-out. Profiler's sniffer_rollup dedupes observed banners into a new
AttackerBehavior.ssh_client_banners JSON column.

Closes gap #3 from SIGNAL_CAPTURE_AUDIT.md.

decnet/profiler/behavioral.py

@@ -90,11 +90,13 @@ def build_behavior_record(events: list[LogEvent]) -> dict[str, Any]:
             _span.set_attribute("tools", ",".join(all_tools))
 
     kex_list = rollup.get("kex_order_raw") or []
+    ssh_banners = rollup.get("ssh_client_banners") or []
     return {
         "os_guess": rollup["os_guess"],
         "hop_distance": rollup["hop_distance"],
         "tcp_fingerprint": json.dumps(rollup["tcp_fingerprint"]),
         "kex_order_raw": json.dumps(kex_list) if kex_list else None,
+        "ssh_client_banners": json.dumps(ssh_banners) if ssh_banners else None,
         "retransmit_count": rollup["retransmit_count"],
         "behavior_class": behavior,
         "beacon_interval_s": beacon_interval_s,

decnet/profiler/fingerprint.py

@@ -21,6 +21,8 @@ _SNIFFER_FLOW_EVENT: str = "tcp_flow_timing"
 _PROBER_TCPFP_EVENT: str = "tcpfp_fingerprint"
 # Prober-emitted HASSHServer fingerprint; carries the raw kex_algorithms string.
 _PROBER_HASSH_EVENT: str = "hassh_fingerprint"
+# Sniffer-emitted SSH client identification string (RFC 4253 §4.2).
+_SNIFFER_SSH_BANNER_EVENT: str = "ssh_client_banner"
 
 # Canonical initial TTL for each coarse OS bucket.  Used to derive hop
 # distance when only the observed TTL is available (prober path).
@@ -75,6 +77,8 @@ def sniffer_rollup(events: list[LogEvent]) -> dict[str, Any]:
     retransmits = 0
     kex_order_raw: list[str] = []
     _kex_seen: set[str] = set()
+    ssh_client_banners: list[str] = []
+    _ssh_banner_seen: set[str] = set()
 
     for e in events:
         if e.event_type == _SNIFFER_SYN_EVENT:
@@ -122,6 +126,15 @@ def sniffer_rollup(events: list[LogEvent]) -> dict[str, Any]:
                 kex_order_raw.append(kex)
                 _kex_seen.add(kex)
 
+        elif e.event_type == _SNIFFER_SSH_BANNER_EVENT:
+            # Sniffer-observed SSH identification string from attacker.
+            # Dedup: the same attacker will reuse the same client banner
+            # across flows/reconnects; record distinct values in order seen.
+            banner = e.fields.get("ssh_version")
+            if banner and banner not in _ssh_banner_seen:
+                ssh_client_banners.append(banner)
+                _ssh_banner_seen.add(banner)
+
         elif e.event_type == _PROBER_TCPFP_EVENT:
             # Active-probe result: prober sent SYN to attacker, got SYN-ACK back.
             # Field names differ from the passive sniffer (different emitter).
@@ -173,4 +186,5 @@ def sniffer_rollup(events: list[LogEvent]) -> dict[str, Any]:
         "tcp_fingerprint": tcp_fp or {},
         "retransmit_count": retransmits,
         "kex_order_raw": kex_order_raw,
+        "ssh_client_banners": ssh_client_banners,
     }

decnet/sniffer/fingerprint.py

@@ -61,9 +61,35 @@ _BUS_TRAFFIC_EVENTS: frozenset[str] = frozenset({
     "tls_session",
     "tcp_flow_timing",
     "tcp_syn_fingerprint",
+    "ssh_client_banner",
 })
 
 
+def _parse_ssh_banner(data: bytes) -> str | None:
+    """
+    Return the attacker's SSH identification string (RFC 4253 §4.2) if
+    *data* begins with one, else None.
+
+    A valid banner starts with ``SSH-`` and terminates at the first CR or LF
+    within the 255-byte RFC-mandated window.  The returned string is decoded
+    as ASCII and stripped of the trailing CR/LF bytes.
+    """
+    if not data.startswith(b"SSH-"):
+        return None
+    end = -1
+    # RFC 4253: identification string (incl. CR LF) must not exceed 255 bytes.
+    for i, b in enumerate(data[:255]):
+        if b in (0x0D, 0x0A):  # CR or LF
+            end = i
+            break
+    if end < 5:  # "SSH-X" minimum
+        return None
+    try:
+        return data[:end].decode("ascii", errors="strict")
+    except UnicodeDecodeError:
+        return None
+
+
 # ─── TCP option extraction for passive fingerprinting ───────────────────────
 
 def _extract_tcp_fingerprint(tcp_options: list) -> dict[str, Any]:
@@ -1053,6 +1079,29 @@ class SnifferEngine:
         if not payload:
             return
 
+        # SSH client banner (RFC 4253 §4.2): attacker→decky TCP/22, first
+        # application-data segment of the flow. Emit once per flow.
+        if (
+            dst_port == 22
+            and dst_ip in self._ip_to_decky
+            and direction_forward
+        ):
+            flow = self._flows.get(flow_key)
+            if flow is not None and not flow.get("ssh_banner_seen"):
+                banner = _parse_ssh_banner(payload)
+                if banner is not None:
+                    flow["ssh_banner_seen"] = True
+                    target_node = self._ip_to_decky[dst_ip]
+                    self._log(
+                        target_node,
+                        "ssh_client_banner",
+                        src_ip=src_ip,
+                        src_port=str(src_port),
+                        dst_ip=dst_ip,
+                        dst_port=str(dst_port),
+                        ssh_version=banner,
+                    )
+
         if payload[0] != _TLS_RECORD_HANDSHAKE:
             return
 

decnet/web/db/models.py

@@ -182,6 +182,13 @@ class AttackerBehavior(SQLModel, table=True):
         default=None,
         sa_column=Column("kex_order_raw", Text, nullable=True),
     )  # JSON list[str] — kex_algorithms comma-separated strings
+    # Sniffer-observed SSH client identification strings (RFC 4253 §4.2),
+    # deduped in observation order. Captures the attacker's SSH client
+    # software (e.g. "SSH-2.0-OpenSSH_9.2p1", "SSH-2.0-libssh2_1.10.0").
+    ssh_client_banners: Optional[str] = Field(
+        default=None,
+        sa_column=Column("ssh_client_banners", Text, nullable=True),
+    )  # JSON list[str]
     retransmit_count: int = Field(default=0)
     # Behavioral (derived by the profiler from log-event timing)
     behavior_class: Optional[str] = None          # beaconing | interactive | scanning | brute_force | slow_scan | mixed | unknown

decnet/web/db/sqlmodel_repo.py

@@ -683,6 +683,16 @@ class SQLModelRepository(BaseRepository):
                 d["kex_order_raw"] = []
         elif raw_kex is None:
             d["kex_order_raw"] = []
+        # Same list-or-None pattern for ssh_client_banners.
+        raw_banners = d.get("ssh_client_banners")
+        if isinstance(raw_banners, str):
+            try:
+                parsed_banners = json.loads(raw_banners)
+                d["ssh_client_banners"] = parsed_banners if isinstance(parsed_banners, list) else [parsed_banners]
+            except (json.JSONDecodeError, TypeError):
+                d["ssh_client_banners"] = []
+        elif raw_banners is None:
+            d["ssh_client_banners"] = []
         return d
 
     @staticmethod