drift

Does a device's real network traffic match what it's declared to be allowed to contact? Checked against a real packet capture.

witness checks a website's own traffic against its declared privacy claims, from inside a browser. drift is the same idea pushed past that ceiling: a device on your own network — anything, not just something with a browser in it — checked against a declared allow-list by reading what it actually sent, from a packet capture.

$ drift check http.cap claims.yaml
[XX] 145.254.160.237: 'test workstation' contacted domain(s) outside its declared allow-list: pagead2.googlesyndication.com

0/1 pass, 1 fail

(Real output — see "Tested against real traffic" below for what capture this is and why.)

Read this before anything else: what's actually verified here

The idea this comes from (see the portfolio audit) is a Raspberry Pi acting as a live network tap on a real home network, checking a real smart-home device's real traffic in real time. That deployment is not built or tested here — this session has no access to a Raspberry Pi, a real IoT device, or a real home network to verify any of it against, and building software nobody can verify would be the one thing this whole portfolio has consistently refused to do.

What genuinely is built and tested: the analysis core -- reading a .pcap/.pcapng file, working out what each device contacted (DNS lookups, TLS SNI, HTTP Host: headers, and IPs it reached without naming), and checking that against a declared allow-list or the manufacturer's own MUD file. That core doesn't care whether the capture came from a live Pi tap, a router's own packet-capture feature, or a file someone handed you -- a .pcap file is a .pcap file. Wiring this to a live tap on real hardware is real, unstarted future work, stated plainly rather than implied by silence.

Scope

Same discipline as witness: this is for checking traffic you captured on a network you own, against claims you declared. It has no notion of any real device manufacturer's actual claims and ships with no data about any real product.

Install

pip install drift-evidence   # the command it installs is `drift`

(drift was already taken on PyPI — same story as most siblings in this portfolio.)

Use

Declare what each device is allowed to contact, in claims.yaml:

devices:
  - id: "192.168.1.42"             # an IP -- or a MAC, "aa:bb:cc:dd:ee:ff",
    name: "smart bulb"             #   which survives a DHCP lease change
    allowed_domains:
      - vendor.example.com         # exactly this domain
      - "*.cdn.example.net"        # any subdomain of cdn.example.net, at any depth
    allowed_ips:                   # direct-IP contacts that are fine (optional)
      - "203.0.113.0/24"
    mud: bulb.mud.json             # the maker's RFC 8520 MUD file (optional)

A plain entry matches only itself: vendor.example.com does not cover telemetry.vendor.example.com. A leading *. covers every subdomain but not the domain itself, so list both if both are allowed. No other wildcard form is accepted, so an allow-list never quietly covers more than it names.

MUD files. RFC 8520 Manufacturer Usage Descriptions are the IETF standard for a manufacturer to publish exactly what its device needs to talk to. Point mud: at one (a path relative to claims.yaml) and every DNS name in its access lists joins the allow-list -- so the check becomes "does this device do only what its maker says it does," in the maker's own words.

Then check a real capture against it:

drift check capture.pcap claims.yaml

A device with no entry in claims.yaml reads unverified, never a silent pass -- same three-status discipline as every claim-checking tool in this portfolio. --json prints the full report. Exit code 1 if any device failed.

Don't have a claims file yet? drift learn writes one from what a capture shows:

drift learn capture.pcap --device 192.168.1.42 --by-mac > claims.yaml

It's a baseline to review, not something to trust blindly: anything the device was already doing wrong during that capture is now in its allow-list. Capture a fresh setup, learn once, then check every later capture against it -- that's what catches a firmware update that starts talking to somewhere new.

How domains are actually extracted

Four signals, none of which needs anything decrypted:

• DNS query names -- the device asking "where is X", visible even when the connection to X itself is encrypted.
• TLS SNI -- the server name a TLS ClientHello sends in the clear before encryption starts. This names almost every HTTPS connection a modern device makes, including ones to a host it resolved some other way (DNS-over-HTTPS, an answer cached from before the capture started). ClientHellos split across TCP segments -- common now that post-quantum key shares make them bigger than one packet -- are reassembled first.
• HTTP Host: headers -- cleartext HTTP only.
• Direct IP contacts -- connections to a public address the device never looked up (no DNS answer anywhere in the capture) and never named (no SNI, no Host:). A device that talks to hardcoded IPs is exactly the one a DNS-only check misses. Each one fails the device unless it's in allowed_ips. Local-network traffic is never counted.

Devices are the addresses that initiated something; a server that only answered never shows up as a device.

What this does NOT do

• Doesn't decrypt QUIC. HTTP/3's ClientHello is encrypted with keys derived from the connection ID -- recoverable, but real work of its own. A QUIC connection to an address the device never resolved over plain DNS therefore shows up as a direct-IP contact, not by name.
• Doesn't see names inside DNS-over-HTTPS -- only the TLS SNI of the DoH server itself.
• No live capture of anything. drift reads a .pcap file that already exists; it has no code that touches a network interface. Making one (a Raspberry Pi tap, a router's mirror port, tcpdump itself) is the caller's job.
• IP identity across a capture. A claims file can name a device by MAC, and drift records the MAC seen for each source IP. But within one capture, if two devices held the same IP at different times, they're still counted as one.

Privacy

Everything stays on the machine drift runs on. It makes no network calls of its own -- it reads the .pcap file from disk, streamed, and the report goes to stdout or a --json file you name; nothing is uploaded or phoned home anywhere.

The thing to actually be careful of is upstream of drift, not in it: a .pcap captured on a shared network (a router's mirror port, a home Wi-Fi capture) records every device's traffic that happened to be on the wire, not just the one you're checking. extract_domains (in pcap.py) buckets by every source IP it sees, and check_pcap reports on every one of them -- a device with no entry in claims.yaml still comes back unverified, with its full list of observed domains included in the report, exactly the same as a declared device's would be. Point drift at a capture that includes a housemate's phone or a guest's laptop, and their browsing domains end up in your output, not just your smart bulb's.

That's not a bug to fix -- filtering out "devices you didn't mean to capture" isn't something drift can know how to do, since it has no way to tell an incidental bystander's IP from a device you meant to declare later. It's a fact about what a shared-network capture is, and worth knowing before you paste a report somewhere or hand it to someone else: capture only what you mean to, and treat the report as covering everyone who was on the network at the time, not just the device named in your claims file.

Tested against real traffic, not synthetic fixtures

tests/fixtures/http.cap is a real capture downloaded from Wireshark's own official sample-captures page (see tests/fixtures/PROVENANCE.md) — 2004-era traffic of a client requesting a page from www.ethereal.com (Wireshark's former name) whose response pulled in a Google AdSense ad from pagead2.googlesyndication.com. That's a real DNS query and a real HTTP request to an undeclared third party, sitting in a 20-year-old capture nobody built for this purpose — exactly the shape drift exists to catch, found in genuine traffic rather than constructed to prove a point.

Also run against captures from Chris Sanders' Practical Packet Analysis (chrissanders/packets):

• cryptowall4_c2.pcapng, a real CryptoWall 4 infection. Declare the host may reach *.microsoft.com and *.windowsupdate.com, and drift check fails it on homealldaylong.com -- the ransomware's command-and-control domain.
• lotsofweb.pcapng (10 MB of browsing) taught drift learn three kinds of junk, all fixed in 0.3.0: googleads.g.doublecl, a Host header cut off where the request split across TCP segments (request heads are now reassembled, and only a complete line counts); 128.16.16.172.in-addr.arpa, a reverse lookup of the device's own address; and digg.com.localdomain-style search-domain retries and .local mDNS names. None of those is a destination.

Tests

pip install -e .
python tests/test_pcap.py      # extraction, against the real fixture above
python tests/test_signals.py   # SNI, direct IPs, MAC ids, allowed_ips, MUD, learn
python tests/test_claims.py    # claims.yaml validation
python tests/test_check.py     # pass/fail/unverified classification
python tests/test_cli.py       # the real CLI entry point, against the real fixture

test_signals.py uses a real ClientHello produced by the local OpenSSL (through Python's ssl module), split across two TCP segments inside a capture generated with scapy -- a real device's traffic can't be committed to a public repo, but the bytes being parsed are real.

MIT licensed.