English · Русский
Wirewolf is an offline desktop co-pilot for malware traffic analysis. Point it at a PCAP from a sandbox run or a perimeter dump and it finds the C2, the payload delivery and the exfiltration, explains why in plain sentences, and takes you to the exact frame — in its own packet viewer or in Wireshark.
No cloud upload, no LLM narratives: every statement is a deterministic rule template filled with measured values and frame numbers.
Existing tools are each strong at one step. Wireshark dissects perfectly but does not tell you what is malicious; NetworkMiner recovers artefacts but does not detect; RITA finds beacons but needs Zeek and Docker and never shows a packet; online sandboxes want your capture. Wirewolf closes the loop find → explain → prove in packets → export IOCs on a single Windows or Linux box, offline.
| Victim identification | IP, MAC, host name, user and AD domain from DHCP, NBNS, NetBIOS datagrams, CLDAP, Kerberos AS-REQ and NTLMSSP (SMB, DCERPC, HTTP) |
| Behavioural detection | beaconing across connections and heartbeats inside one long session, DGA, DNS tunnelling, payload delivery by magic bytes vs Content-Type, TLS certificate anomalies, custom protocols on 443/80, recon → upload chains, XLoader-style domain rotation, non-browser check-ins, stale hard-coded user agents, scanning, exfil ratios, inbound web exploitation |
| Verdict model | Noise · Unknown · Suspicious · Likely · Confirmed; one signal never goes above Suspicious; risk and confidence are separate; every finding lists what speaks against it |
| Fingerprints | JA3/JA3S, JA4, JA4S, JA4H, JA4X, JA4T, QUIC Initial decryption (v1, v2, drafts) for SNI and JA4 |
| Decryption hub | SSLKEYLOGFILE or pcapng-embedded secrets → decrypted HTTP/1.1 and HTTP/2 metadata feeds the same detectors |
| Stream transforms | base64 · hex · URL · gzip/deflate · HTTP body · XOR with 1–4 byte key guessing, applied to any TCP stream |
| Packet viewer | virtualised list of millions of frames, TShark field tree, hex view with the bytes behind each verdict highlighted, Wireshark display filters |
| Wireshark, both ways | open any finding at the right frame with the right filter; a Lua post-dissector adds wirewolf.verdict, wirewolf.risk, wirewolf.finding fields inside Wireshark |
| Reports & IOCs | Markdown incident report, defanged IOC table, CSV, local threat-intel feeds (abuse.ch lists, your own) |
| Built for hostile input | streaming parser in an isolated worker process with timeouts, per-frame panic containment, mutation fuzzing; nothing from the traffic is ever written to disk or executed |
Wirewolf is developed against the complete public archive of malware-traffic-analysis.net — training exercises with answer keys and years of blog posts. The current snapshot (full report):
| Captures parsed | 771 PCAPs (≈19 GB), 0 crashes, 0 timeouts |
| Agreement with TShark | frames 100 %, tcp.stream 99.9 % (±1 %), udp.stream 100 %, SNI 94 %, JA4 93 % |
| Exercises with answer keys | victim IP 87.5 %, user account 87 %, a C2 from the answers in the top-5 threats 90 % |
| Held-out years (2020+) | victim IP 93 %, C2 in top-5 100 % |
| Blog infection captures | 95 % produce a Suspicious+ finding, 59 % a Likely one |
| Throughput | 50–95 MB/s including database writes, worker peak memory ≤ 410 MB |
Detectors are tuned on years ≤ 2019 and checked on later years so rules generalise instead of memorising IOCs. Answer keys are extracted automatically from PDFs and contain errors; see the report for a per-file breakdown.
Prerequisites: Rust stable, Wireshark ≥ 4.2 (TShark is used for the field tree, decryption and as a test oracle), Node.js ≥ 20 for the UI.
git clone https://github.com/grabskiyr/wirewolf.git
cd wirewolf
cargo build --release # ww (CLI + isolated parsing worker)
cd app && npm install
npx tauri build # Windows installer in target/release/bundle/During development, npx tauri dev opens the window with hot reload. The same UI also runs in a browser on top of a local API:
target/release/ww serve --ui app/dist # http://127.0.0.1:8765, localhost onlyww analyze capture.pcap # victim, threats, Wireshark commands
ww analyze capture.pcap --keylog sslkeys.txt # decrypt TLS first
ww analyze capture.pcap --report md # capture.pcap.wirewolf.md
ww analyze capture.pcap --wireshark # verdict table for the Lua plugin
ww analyze capture.pcap --json # summary + detection as JSON
ww stream capture.pcap --stream 14 --server --recipe httpbody,base64,gunzip
ww wireshark capture.pcap --filter 'wirewolf.verdict == "Likely"' --frame 1234
ww eval corpus/ --oracle # corpus regression vs TSharkKeys in the packet viewer: N / P next / previous evidence packet · G go to frame · S packets of this stream · W open in Wireshark · Alt+1…6 switch view.
ingest ─▶ parse ─▶ flows / sessions ─▶ identity ─▶ signals ─▶ findings ─▶ UI · reports · Wireshark
pcap L2–L7 TShark-aligned DHCP…NTLM rules verdicts
pcapng tunnels stream numbers ATT&CK victim
gz/zst/zip QUIC frames
└──────────── DuckDB per capture, cached by SHA-256 ────────────┘
| Crate | Role |
|---|---|
ww-core |
data model shared by everything |
ww-parse |
streaming PCAP/PCAPNG reader, decapsulation (VLAN, MPLS, PPPoE, GRE, ERSPAN, VXLAN, GTP-U, IP-in-IP, Teredo), defragmentation, TCP reassembly, DNS, TLS, QUIC, HTTP, host identity, JA3/JA4+ |
ww-store |
DuckDB schema, streaming appenders, parse cache |
ww-detect |
detectors, scoring, verdicts, victim ranking, local feeds |
ww-decrypt |
key logs via TShark, stream transforms |
ww-report |
Markdown report, Wireshark verdict table |
ww-wireshark |
TShark/Wireshark discovery and the test oracle |
ww-api |
one JSON dispatcher for the UI (Tauri commands and ww serve) |
ww-cli |
ww binary: analyze, eval, stream, serve, worker |
app/ |
Tauri 2 + React + TypeScript + Tailwind UI |
plugins/wireshark/wirewolf.lua |
Wireshark post-dissector |
Captures of malware traffic contain live samples. Wirewolf reads packets as data: it never runs anything, never exports objects to disk and never contacts the network unless you explicitly run scripts/update_feeds.py. The corpus tooling refuses to download sample archives and only unpacks members with a PCAP/PCAPNG signature. Still, analyse untrusted captures inside a VM. See SECURITY.md.
- Streaming parser, TShark-aligned streams, JA4+ — stage 0
- Detectors, verdict model, victim identification — stage 1
- Desktop UI with packet viewer and Wireshark integration — stage 2
- TLS decryption from key logs, stream transforms — stage 3 (part)
- Kerberos / NTLM / WPA / IPsec decryption via the hub
- Network graph, investigation board, automatic attack chains
- Baseline diff against clean VM captures, Zeek/Suricata plugins, retro-hunt
- STIX 2.1 / MISP export, opt-in VirusTotal/MISP lookups
The full specification (in Russian) is in docs/SPEC.md.
Dual-licensed under MIT or Apache-2.0, at your option. JA4+ methods other than JA4 itself are © FoxIO and covered by the FoxIO License 1.1 (research and internal use). Test traffic comes from malware-traffic-analysis.net and is not redistributed here.




