Build plan · revised 2026-11-10
A repeatable way to learn what every device on a home and lab network collects, what it sends out, and whether the two are linked. Each audit ends as a firewall allowlist.
The starting point is The Homeboy Workshop's tap of a rooted LG OLED. The TV sat on an isolated segment behind an inline hardware tap. A laptop answered DNS and HTTP as a fake internet, with its own certificate authority, and advertised fake devices. The numbers below come from the video's transcript and description.
Handshakes by outcome, fake internet with a private CA
Unicast requests to canaries, "in a few hours"
One hour of events, drawn from the rates the video gives
| Gap in the video | Change here |
|---|---|
| One network condition. The fake internet stops most conversations at the certificate. | Three conditions: real internet, fake internet, no answer. |
| Collection is shown. Exfiltration is not. | Unique canary tokens, plus repeated A/B upload comparison. |
| Treats the server's private key as the route to plaintext. Most TLS 1.2 is also ephemeral, and the vendor's key is never available. | Plaintext at the endpoint: device key log, a proxy the device accepts, or metadata only. |
| One capture, one device state. | A scripted state matrix. |
| No radio visibility, so the Wi-Fi scan claim stayed open. No mention of IPv6. | Radio sniffers for management frames, BLE and 802.15.4. IPv6 captured and attributed by MAC. |
| One inline tap per device. | A whole-home sensor first. The tap rig becomes the second tier. |
No single capture point answers all three questions. A capture on the router trunk sees everything that is routed, but not two devices talking inside one VLAN on one switch. The plan combines five sources.
| Behaviour | Trunk capture | Resolver logs | Canaries | Port mirror or bench tap | Radio sniffer | Host agent |
|---|
| Layer | Field | Readable | Limit |
|---|---|---|---|
| 2 | MAC addresses, VLAN tag | Yes | Identifies the device on its own segment. |
| 3 | IP addresses | Yes | IPv6 privacy addresses rotate. Attribute by MAC. |
| 4 | Ports, packet sizes, timing | Yes | Enough for volumes and beacon periods. |
| 7 | DNS names | Yes | Not when the device uses DNS over HTTPS or TLS. |
| TLS | Server name, client fingerprint | Yes | Encrypted Client Hello hides the server name. |
| TLS | Server certificate | TLS 1.2 only | TLS 1.3 encrypts it. |
| 7 | Payload | No | Needs keys from the device, or a proxy the device trusts. |
Tier 1 always on
Tier 1 watches every device cheaply and all the time. It answers question 2 in metadata, catches discovery and probing, and tells you which devices deserve the bench.
Phase 0
Done whenEvery MAC seen in a 7-day lease log maps to a row.
/ip dhcp-server lease. Neighbours: /ip arp, /ipv6 neighbor./interface bridge host.Phase 1
trusted, iot, lab, guest, mgmt, bench.Done whenResolver logs show per-device queries, and a device with a hard-coded resolver shows up in the redirect counters.
dst-nat rules in /ip firewall nat for UDP and TCP 53, with the resolver's own address excluded./ipv6 firewall. Repeat the redirect and blocks there./ip upnp. While it is on, its mappings appear as dynamic NAT rules.iot and bench networks during an audit.Phase 2
Capture the trunk between the router's LAN side and the core switch. This is before NAT, so every flow keeps its device MAC and address. Use a mirror port or one inline tap.
Sensor capture interfaces carry no IP address, emit nothing, and keep the VLAN tag:
ip link set eth1 promisc on arp off up
sysctl -w net.ipv6.conf.eth1.disable_ipv6=1
ethtool -K eth1 gro off lro off gso off tso off rx off tx off rxvlan off
Capture into a ring. A new file starts every 300 s or every 1 GB, whichever comes first:
dumpcap -i eth1 -i eth2 -w /data/pcap/trunk.pcapng \
-b duration:300 -b filesize:1000000 -b files:2000
iot, lab and bench. For trusted and guest, keep Zeek metadata and delete the packets.Disk in GB = average Mb/s × days × 10.8
| Average | 3 days | 7 days | 14 days |
|---|---|---|---|
| 1 Mb/s | 32 GB | 76 GB | 151 GB |
| 5 Mb/s | 162 GB | 378 GB | 756 GB |
| 20 Mb/s | 648 GB | 1.5 TB | 3.0 TB |
| 50 Mb/s | 1.6 TB | 3.8 TB | 7.6 TB |
Done whenA test flow from a known device appears in the capture with the right MAC and VLAN tag, in both directions, over IPv4 and IPv6.
/interface ethernet switch. Mirror the router trunk port to the sensor port. Property names differ by RouterOS version and switch chip./tool sniffer can stream packets to the sensor as TZSP, and /ip traffic-flow exports NetFlow or IPFIX. Both see only packets that reach the CPU. Hardware-offloaded switching, and layer-3 hardware offload on CRS models, bypass them. Use the mirror for evidence and these for quick looks.horizon.tcpdump over SSH. Recent versions can also export flows as NetFlow or IPFIX. The built-in views use the vendor's own classification, so use the mirror for evidence.Phase 2, wireless
Done whenA Wi-Fi test device's flows appear on the trunk capture, and a second Wi-Fi device cannot reach it directly.
iot VLAN, client isolation on, 2.4 GHz and 5 GHz only, multi-link operation off for that SSID.mgmt. If a controller or cloud service manages them, audit that traffic too.tcpdump on the access point can show client-to-client traffic that never reaches the wire.Phase 3
Canaries are fake devices in each VLAN. Anything that contacts one is enumerating the network.
Token rule. Every canary name, MAC, serial and model string contains a unique random token, such as printer-k7q2m9xw. One token per canary per VLAN. A later sighting of a token in decrypted traffic, a vendor privacy export or an ad-profile page is proof of exfiltration.
Done whenA port scan and an mDNS browse from a laptop both show up in the canary log.
Phase 4
Zeek turns packets into logs. Load MAC and VLAN logging so every connection carries its device and segment:
zeek -C -r trunk_00001.pcapng \
policy/protocols/conn/mac-logging policy/protocols/conn/vlan-logging \
LogAscii::use_json=T
Keep the logs long-term and the packets for days. Query the logs directly:
SELECT orig_l2_addr, "id.resp_h", sum(orig_bytes) AS up
FROM read_json_auto('conn*.log')
GROUP BY 1, 2 ORDER BY up DESC;
Standing queries per device
Done whenThe standing queries run from one command and produce a per-device table.
Phase 5
Run from each VLAN, monthly and after firmware updates:
nmap -sS -sU --top-ports 200 -sV -oA scan-v4 <target>
nmap -6 -sS -sV -oA scan-v6 <target-v6>
/ip neighbor discovery-settings./ip cloud, /ip service, the MAC server and RoMON settings, and record what each device contacts.mgmt and give it a scorecard.Tier 2 one device at a time
The bench is the video's rig with the gaps closed. Use it for devices tier 1 flags and for new purchases before they join the network. Bench first the devices that score highest:
Capture hours per cell
| Device state | Real internet | Fake internet | No answer |
|---|---|---|---|
| Out of box, terms not accepted | 2 h | 2 h | 2 h |
| Terms accepted, all opt-outs | 2 h | 2 h | 2 h |
| Terms accepted, all opt-ins | 2 h | 2 h | 2 h |
| Standby | 24 h | 24 h | 24 h |
| Active use | 2 h | 2 h | 2 h |
| External input playing | 2 h | 2 h | 2 h |
| After firmware update | 2 h | 2 h | 2 h |
| Column total | 36 h | 36 h | 36 h |
One scorecard per device. The example below fills the scorecard from the video alone. The open rows are what the extra conditions, tokens and radio capture exist to answer.
Turn the observed destinations and ports into a per-device firewall allowlist. This is the idea behind Manufacturer Usage Description (RFC 8520). Alert on any flow outside the list, and re-audit after each firmware update.
On servers, VMs and containers you control, the host tells you more than the wire: per-process egress attribution, TLS key logs, and container egress through one logged gateway. Keep wire capture for opaque boxes such as management controllers, NAS units, cameras and recorders.
iot, lab and bench.bench.