Maltrail

Maltrail is a network traffic detection system that identifies communication with known malicious infrastructure and reports selected traffic anomalies. It matches domains, URLs, IP addresses, IP:port pairs, and User-Agent values observed on the network against a set of indicators called trails.

A detection is recorded as a single event containing the source, destination, protocol, matched trail, classification, and trail source:

"2026-08-07 09:14:22.117034" gw 10.13.13.2 57809 1.1.1.1 53 UDP DNS malware.bakewithdavid.com "asyncrat (malware)" (static)

Maltrail is designed for indicator-based network monitoring. Its heuristic detections supplement trail matching, but it is not a replacement for endpoint telemetry or a general-purpose intrusion prevention system.

Features

• A full trail build combining more than 3,000 bundled static files, 42 public-feed integrations, and optional operator-supplied trails.
• A multithreaded Rust sensor using libpcap, with optional Linux PACKET_FANOUT capture workers.
• A Python server providing the reporting interface, event intake, and HTTP API.
• Plain-text custom trails and whitelists that can be reviewed and version-controlled.
• Heuristics for scanning, DNS exhaustion, DGA-like lookups, suspicious downloads, proxy probes, suspicious User-Agent values, and related network activity.
• Local event logging, remote Maltrail logging, CEF over syslog, and Logstash JSON output.
• Deployment validation with maltrail-sensor -T and optional Prometheus metrics.

Contents

• Architecture
• Reporting interface
• Performance
• Installation
  • Installer
  • Building from source
  • Systemd
  • Docker
• Configuration
• Trails
• Events and API
• Operations
  • Monitoring
  • Event retention
• Documentation
• Contributing
• Project
  • License
  • Maintainers
  • Sponsors
  • Presentations and publications
  • Derived blacklist
  • Third-party integrations
  • Acknowledgements

Architecture

Maltrail consists of two independent processes that may run on the same host or on separate hosts:

   ┌──────────┐   events (UDP or file)   ┌──────────┐
   │  sensor  │ ───────────────────────► │  server  │ ◄── browser
   └──────────┘                          └──────────┘
    Rust                                  Python
    libpcap + PACKET_FANOUT               reporting UI + API
    trail matching + heuristics

The sensor captures traffic, performs trail matching and heuristic analysis, and produces events. It can write events locally (LOG_DIR), send them to a remote Maltrail server (LOG_SERVER), or do both. It can also emit CEF over syslog (SYSLOG_SERVER) and JSON to Logstash (LOGSTASH_SERVER).

The server receives and stores remote events, serves locally available event logs, and provides the web interface and API.

Reporting interface

Maltrail includes a browser-based reporting interface for exploring detected traffic, with live updates, field-aware search, retro hunting, geographic views, triage, saved views and export.

The interface is served by server.py at HTTP_ADDRESS:HTTP_PORT. It is plain JavaScript with a single third-party runtime dependency (PapaParse, for CSV parsing) and no build step. One day is viewed at a time, selected with a date picker that doubles as an event-density grid over the available daily logs. Events are streamed from /events and aggregated in the browser into threats — one row per distinct (source, trail) — shown in a sortable grid with a detail panel.

Triage state, saved views, tags and appearance settings are stored in the browser (localStorage), not on the server: they are per-browser and per-origin, and are not shared between analysts.

Sessions restricted with a network filter see only events from their own networks, and that restriction applies to the counts, map and blacklist endpoints as well as to the event list.

Country and ASN enrichment for individual addresses is looked up at stat.ripe.net by the server, which caches the results and serves them to the interface from its own /ripe endpoint; the browser talks to nothing but Maltrail. Set DISABLE_RIPE_LOOKUPS to switch the outbound lookups off entirely. Without them — or on a host with no internet access — flags come from the local RIR table instead and everything else in the interface works offline.

Performance

Performance depends on processor, traffic composition, trail-set size, capture driver, and network interface. The figures below measure the sensor's packet-processing path in isolation; they are not end-to-end live-capture measurements.

Representative measurements on an AMD Ryzen 7 PRO 4750U with heuristics enabled and a 1.5 million-row trail set:

Offline comparison runs using the same generated capture, configuration, and trail set measured a 14–37× lower steady-state per-packet cost than the retired Python sensor across the tested systems. Those figures separate whole-process time from steady state, because trail loading dominates a short replay. Detection itself is asserted separately, by the 42-case corpus in sensor/tests/replay.rs.

Measure it on the target system with:

cargo bench --manifest-path sensor/Cargo.toml --bench hotpath

One capture worker is used by default. Additional workers can increase capture capacity, but Linux flow hashing divides per-source state between workers and therefore reduces the sensitivity of some scan heuristics. In the documented test, 91% of single-worker heuristic alerts remained with two workers, 86% with four, and 65% with eight. Exact trail matching was unchanged. Increase CAPTURE_FANOUT only when capture-drop metrics show that it is necessary.

Benchmark methodology, hardware results, profiler output, memory measurements, and live fanout checks are documented in sensor/docs/REPORT.md.

Installation

Installer

The installer is verified on twelve Linux distributions — Debian, Ubuntu, Fedora, Rocky, AlmaLinux, Arch, openSUSE Leap and Tumbleweed, and Alpine — plus FreeBSD and macOS, on every release, with the full result recorded in docs/compat. Raspberry Pi OS and other 64-bit ARM systems use the aarch64 build; 32-bit ARM has no prebuilt sensor and must build it from source.

curl -fsSL https://raw.githubusercontent.com/stamparm/maltrail/master/install.sh | sudo sh

It installs dependencies, creates a managed checkout under /opt/maltrail, verifies the checksum of the prebuilt sensor, creates an unprivileged maltrail account, installs systemd units, prepares the log and state directories, and starts the sensor and server. Re-running the installer upgrades the managed checkout.

Review the script before running it with elevated privileges. From an existing checkout, the dry run shows the commands without changing the system:

sh install.sh --dry-run

Common installer options:

sh install.sh --role sensor      # Install only the sensor
sh install.sh --ref 3.1.2        # Install a release tag instead of master
sh install.sh --no-service       # Install without changing systemd
sh install.sh --dry-run          # Print commands without applying them
sh install.sh --uninstall        # Remove the managed installation; keep logs and state

The dashboard is available at http://127.0.0.1:8338 after installation. Note that the shipped HTTP_ADDRESS is 0.0.0.0, so it is reachable on every interface, not only loopback — and the default credentials are admin / changeme!. Change USERS, and set HTTP_ADDRESS to 127.0.0.1 (or put the server behind a reverse proxy with TLS), before the host is on an untrusted network.

The initial trail build can take several minutes. The sensor does not detect trail matches until a valid trail set is available. The systemd unit runs the sensor's -T validation before startup so that missing privileges, an unwritable log directory, or an invalid trail set causes startup to fail visibly.

The installer test harness covers twelve distributions, and "it installed" is not the assertion: in each one the server is started and asked for /ping, the sensor is asked to validate itself with -T, the units are checked for paths that resolve, the installer is re-run to prove an upgrade keeps operator configuration, and --uninstall is run. Every result is recorded per platform in docs/compat, and the page there is generated from those rows rather than written by hand.

Alpine and other musl systems get a -musl sensor build. They used to be told the prebuilt binary was glibc-linked and to compile their own; the sensor builds and runs on musl natively, so that was a missing artefact rather than a platform limit.

Building from source

The sensor requires Rust 1.74 or newer, libpcap development headers, and the system's capability tools. The server and trail updater require Python 3.6 or newer.

Install the distribution packages:

# Debian / Ubuntu / Raspberry Pi OS
sudo apt-get install cargo libpcap-dev libcap2-bin python3

# RHEL / Fedora
sudo dnf install cargo libpcap-devel libcap python3

# openSUSE / SLES
sudo zypper install cargo rust libpcap-devel libcap-progs python311

Then build and validate the sensor:

git clone --depth 1 https://github.com/stamparm/maltrail.git
cd maltrail

cargo build --release --manifest-path sensor/Cargo.toml

sudo setcap cap_net_raw,cap_net_admin=eip \
  sensor/target/release/maltrail-sensor

sudo install -d -o "$USER" -g "$(id -gn)" -m 750 /var/log/maltrail

sensor/target/release/maltrail-sensor -T
sensor/target/release/maltrail-sensor

Start the server in another terminal or on another host:

python3 server.py

Prebuilt sensor binaries are attached to current releases with SHA-256 checksums: Linux x86_64 and aarch64 against both glibc and musl, macOS on Apple silicon and Intel, FreeBSD amd64, and Windows x86_64.

The glibc builds link libpcap statically and target glibc 2.28, so the C library is the only thing they need — nothing to install, on RHEL 8+, Debian 10+, Ubuntu 18.04+ and Leap 15.x alike. The musl builds are fully static, so Alpine needs nothing at all. The Windows build is 64-bit and needs Windows 10 or later plus Npcap installed before it will start — wpcap.dll is a load-time dependency, so without it the loader refuses the executable rather than failing at capture. The archive says so too.

Binaries from 3.1.1 and earlier did not: they linked libpcap dynamically, and asked for it by the name their AlmaLinux build host uses. Debian and Ubuntu ship the identical library under the older name libpcap.so.0.8, so those binaries stop before they start —

./maltrail-sensor: error while loading shared libraries: libpcap.so.1: cannot open shared object file

— on a machine that has libpcap installed. install.sh links the missing name for you. By hand:

# adjust the directory for your architecture: aarch64-linux-gnu, or /usr/lib64 on RPM distributions
sudo ln -sf /usr/lib/x86_64-linux-gnu/libpcap.so.0.8 /usr/lib/x86_64-linux-gnu/libpcap.so.1
sudo ldconfig

Systemd

The supplied packaging/systemd/ units run both processes as the unprivileged maltrail user. Systemd creates /var/log/maltrail and /var/lib/maltrail, restricts filesystem access, and grants the sensor CAP_NET_RAW and CAP_NET_ADMIN.

The installer configures these units automatically. For an existing source installation, follow the manual service procedure in sensor/docs/INSTALL.md.

Check service state and logs with:

systemctl status maltrail-sensor maltrail-server
journalctl -u maltrail-sensor -f

Docker

Start the supplied Compose deployment with:

docker compose -f docker/docker-compose.yml up -d

Container configuration, storage, privileges, and health checks are documented in docker/README.md.

Configuration

Maltrail reads maltrail.conf, which contains separate [Sensor] and [Server] settings. The installer places the managed configuration at /etc/maltrail.conf.

Frequently used sensor options include:

PROCESS_COUNT applies to the retired Python sensor and to the legacy event-log throttle; it does not set the Rust sensor's worker count. Configure capture workers with CAPTURE_FANOUT or CAPTURE_WORKERS instead.

Run the deployment check after changing configuration:

sensor/target/release/maltrail-sensor -T

The check validates configuration, trails, whitelist entries, capture filter, privileges, log storage, update support, and worker settings. A successful check includes positive trail and whitelist counts rather than only confirming that files exist.

Trails

A trail is one indicator — a domain, URL, IP address, IP:port pair, User-Agent, JA3/JA4 fingerprint or certificate hash — together with what it means and where it came from. The updater merges four sources into TRAILS_FILE, in this order:

The static trails live in their own repository. Detection content changes tens of times a day; the engine does not, and keeping them together meant updating detection required pulling code and made this repository's history unusable. STATIC_TRAILS_URL points at the newest published set:

STATIC_TRAILS_URL https://github.com/stamparm/trails/releases/latest/download/trails.csv.gz

Point it at a specific content-YYYYMMDD-HHMM release instead to pin a version, so a bad publish is not immediately global. The set is cached next to TRAILS_FILE, which is what makes an offline or air-gapped rebuild work; the published sha256 is checked before downloading, so a deployment that updates more often than the content changes transfers 65 bytes rather than 11 MB, and a payload that does not match its digest is refused in favour of the cache.

update_trails() publishes a new TRAILS_FILE atomically and only after a successful build. Feeds that return nothing are reported by name, so a deployment does not silently depend on a source that has quietly retired.

Add your own indicators under CUSTOM_TRAILS_DIR, and anything that must never raise an event to USER_WHITELIST. Keep both outside the installation directory so an upgrade cannot overwrite them.

Static trail contributions go to stamparm/trails; new feeds go here. Either way an indicator needs a classification and a source somebody can check — see Contributing.

Events and API

Maltrail records one whitespace-separated event per detection, using CSV quoting where a value contains spaces:

"<time>" <sensor> <src_ip> <src_port> <dst_ip> <dst_port> <proto> <type> <trail> "<info>" <reference>

The type field identifies what matched, including DNS, IP, IPORT, URL, PATH, HTTP, UA, PORT, CERT, JA3, and JA4. The info field contains the trail classification, and reference identifies the static list, feed, custom source, or heuristic that produced it. The JA3/JA4 types fire on TLS client fingerprints: an implant's TLS stack survives every address and domain rotation, so its hello hash keeps matching after everything else has burned (published by the abuse.ch SSLBL JA3 feed).

Indicator lookup

Use /check to query one domain, IP address, or URL:

curl 'http://127.0.0.1:8338/check?q=www.sub.evil.example'

{
  "query": "www.sub.evil.example",
  "found": true,
  "trail": "evil.example",
  "info": "asyncrat (malware)",
  "reference": "(static)",
  "confidence": 100
}

The confidence field (0-100, or null when unavailable) says how strongly the sources back the listing: 40 for a single feed, +15 per additional feed agreeing independently up to 100, and full marks for the operator's own custom and static entries. It is computed at trail-update time from feed agreement into a trails.confidence sidecar next to trails.csv; a server pulling trails from an UPDATE_SERVER has no provenance to score and reports null. Use it to prioritize triage - a single-feed listing at 40 deserves a second look before it earns a firewall rule.

A subdomain lookup can match its listed parent. URL lookups check host/path before checking the host alone. The server reads the memory-mapped trail database and observes trail updates without a restart.

Public static and feed trails are available without authentication, consistent with the /trails endpoint used by remote sensors. Custom trails require an authorized session; an unauthorized custom-only lookup is reported as a miss. Event data remains authenticated.

Operations

Monitoring

Use maltrail-sensor -T as a deployment and configuration gate. The supplied systemd unit runs it as ExecStartPre.

To confirm that detection itself works — not just that the processes start — run:

python3 server.py --detect-test

It replays a crafted pcap of emulated malicious traffic (trail hits on a DNS query, an IP, an IP:port, a URL path and a Host header, plus the SQL-injection, traversal, RCE, XSS, proxy-probe, sinkhole, missing-Host and port/web/infection-scanning heuristics) through the installed sensor and asserts that every expected detection fires. It needs no root, no interface and no trail set of its own. A healthy install prints 20/20 detection(s) fired.

When STATS_ADDRESS is configured, monitor at least these Prometheus metrics:

State saturation affects the corresponding heuristic; exact trail matching remains active.

Send SIGHUP or use systemctl reload maltrail-sensor to request a trail reload. Trail files updated by another process are detected automatically and published to workers without restarting the sensor.

The condensed observable store (USE_CONDENSED_STORAGE, meta.sqlite) supports the server's novelty and retro-hunt views. The per-day event-log sidecar index (USE_EVENT_INDEX, LOG_DIR/index/*.sqlite, roughly twice the log size on disk) is what makes /counts exact and /hunt fast; it is maintained incrementally from the logs themselves and can be rebuilt with server.py --rebuild-index. Compatibility with the retired sensor is documented in sensor/docs/COMPATIBILITY.md.

Event retention

Maltrail does not rotate or delete event logs. Operators are responsible for defining retention, archival, and deletion according to storage requirements and organizational policy.

Recommended practices:

• Send the durable event copy to a remote Maltrail server or SIEM with LOG_SERVER, SYSLOG_SERVER, or LOGSTASH_SERVER.
• Alert on maltrail_log_dir_free_bytes with enough headroom for the expected event rate.
• Rotate, archive, or remove local daily logs using external tooling.
• Keep files needed by the reporting interface uncompressed in LOG_DIR; archive compressed files elsewhere.

When the log filesystem is full, the sensor cannot append events. Event logs may also contain IP addresses and domains that are regulated as personal data in some jurisdictions; retention policy should account for the applicable requirements.

Synthetic traffic

To check that detection and the dashboard both still work, without waiting for real traffic:

python3 server.py --detect-test                      # assert every detection fires, then exit
python3 server.py --detect-test --keep DIR --serve   # ...and keep the events, serving them on :8338

--keep also replays sensor/tests/corpus/ into the same log and prints which of the shapes the dashboard renders differently have an event behind them, so a missing icon, colour or glyph is visible rather than assumed. Timestamps are shifted so the newest day is today. A sensor binary is required (cargo build --release --manifest-path sensor/Cargo.toml).

The public demo's data is regenerated from such a run:

python3 sensor/tools/gen_demo_js.py --from DIR/logs   # tops up html/js/demo.js

Documentation

Contributing

Trail additions, feed maintenance, bug reports, documentation, and sensor improvements are welcome. Trail submissions should include a reliable source and should use the narrowest appropriate classification.

Run the relevant checks before submitting code. The complete sensor gate is:

bash sensor/tools/check.sh

It runs formatting, Clippy with warnings denied, and the debug and release test suites. Run the Python server suite with:

bash tests/run.sh python3

The Windows build can be exercised from Linux, which is where its bugs have been found:

sh sensor/tools/check_windows.sh

It cross-compiles the sensor with mingw-w64, extracts Npcap's userspace library from its installer (an NSIS archive, so nothing is installed), and runs the result under Wine — the whole unit suite, -T against the shipped configuration, the pcap corpus compared byte for byte against the native binary, and the server answering /ping under a Windows Python. Live capture is the one thing it cannot cover; that needs Npcap's kernel driver and a real Windows machine. Prerequisites are gcc-mingw-w64-x86-64, wine, and p7zip-full.

Project

License

TL;DR: Maltrail is MIT-licensed, but the Maltrail Trails dataset has separate terms. Independent IOC lookup/reference is fine; systematic use of Trails as an intelligence source in a commercial product or service requires permission/licensing.

Maltrail is distributed under the MIT License. See LICENSE.

That is the engine. The static trail set is separate work under separate terms: free for internal defensive use, research and teaching, but a commercial product, service, MSSP or MDR offering, or a redistributed feed needs a licence. An MIT engine does not make the content free to sell — see LICENSE.md in stamparm/trails before shipping it in something you charge for.

Maintainers

• Miroslav Stampar (@stamparm)
• Mikhail Kasimov (@MikhailKasimov)

Sponsors

• Sansec (2024–2025)
• Sansec (2020–2021)

Presentations and publications

• 47th TF-CSIRT Meeting, Prague, 2016 (slides)
• Detect attacks on your network with Maltrail, Linux Magazine, 2022 (article)
• Best Cyber Threat Intelligence Feeds, Silent Push, 2022 (review)
• Research on Network Malicious Traffic Detection System Based on Maltrail, Nanotechnology Perceptions, 2024 (paper)

Third-party integrations

• FreeBSD Port
• OPNsense Gateway Plugin
• D4 Project
• BlackArch Linux
• Validin
• Maltrail Add-on for Splunk
• Maltrail decoder and rules for Wazuh
• GScan (trails only)
• MalwareWorld (trails only)
• oisd domain blocklist (trails only)
• NextDNS (trails only)
• NoTracking (trails only)
• OWASP Mobile Audit (trails only)
• Mobile Security Framework MobSF (trails only)
• pfBlockerNG-devel (trails only)
• Sansec eComscan (trails only)
• Palo Alto Networks Cortex XSOAR (trail connector)

Acknowledgements

• Thomas Kristner
• Eduardo Arcusa Les
• James Lay
• Ladislav Baco (@laciKE)
• John Kristoff (@jtkdpu)
• Michael Münz (@mimugmail)
• David Brush
• @Godwottery
• Chris Wild (@briskets)
• Keith Irwin (@ki9us)
• Simon Szustkowski (@simonszu)