Methodology

This page explains the source of the numbers on HoneyLabs, how a connection becomes a report, and what is removed before anything is published. The evidence behind a verdict is on the report itself. If one looks wrong, mail info@honeylabs.net.

The data

HoneyLabs runs honeypot sensors on the public internet. No service is advertised on them, so every connection they receive is unsolicited: a scan, a probe, a password attempt or an exploit. First-contact sensors record the first bytes of each connection. Interactive sensors answer as the real service would, so the client carries on. It logs in, runs commands and fetches a payload, and the sensor records the session. All of it comes from our own sensors, and no third-party feed is mixed in.

For each connection we keep the source IP, the destination port, the probe payload, the protocol fingerprints (JA4, JA3, Akin, HASSH), the user agent when the client sends it, and a UTC timestamp.

From connection to report

Each source IP gets a country and a network from MaxMind GeoLite2, its reverse DNS, and a scanner tag when its reverse DNS and operator match a known research, commercial, hosting or VPN scanner. Payloads are matched against CVE detection patterns from the Nuclei templates, which are refreshed daily and checked against CISA's Known Exploited Vulnerabilities catalog.

A CVE's counts cover both kinds of sensor. A first-contact request counts when it matches a pattern of that CVE and the pattern belongs to no other CVE. A request to an interactive sensor counts when the rule for that CVE is the only rule it matched, or when it carried the CVE's exploit. A request that matched the rules of several CVEs counts for none of them. An address seen by both kinds of sensor is counted once.

The results are stored in ClickHouse and aggregated by IP, port, network, fingerprint and CVE. Lookup, the MCP server and the HTTP API all read those aggregates.

The verdict

The verdict on a report comes from fixed rules over the evidence, and no model scores it. The rules read four things, in this order.

Scanner tags
An IP that belongs to a recognised research or commercial scanner, by reverse DNS and operator, is labelled as that scanner and is not called a threat.
Activity
Volume, the number of ports and how long the IP has been active. Ten or more distinct destination ports is classed as scanning, and fewer than ten as probing.
Exploit signatures
A payload that matches a CVE pattern, or carries a loader or command-injection command such as a piped wget or busybox, raises the verdict to serious or critical. The public exploiter feed is decided by the same evidence, so a report and the feed agree about an address.
Engagement
How many bytes the source sent once something answered, and whether it held sessions open. It is shown beside the verdict and does not change it, because it tells an address that touched a port from one that stayed and talked.

Confidence says how much evidence agrees: a single packet is low, and a stable fingerprint across thousands of events is high. Every input is on the report, so you can check the verdict against the records.

Redaction

Three things are removed or changed in the database before anything is published.

  • Sensor addresses and sensor identities are removed in every encoding, including where they appear inside payloads.
  • Payloads from the first-contact sensors are reduced to the searchable shape of the probe. Logins tried against the interactive sensors are shown as sent, with username and password. Where the client never sent the password, the record says so.
  • Indicators in shared text are defanged, so a report never holds a working link to attacker infrastructure.

Coverage

The sensors see a sample of internet scanning. They are not on every network, so an IP that never reached one may still be hostile, and its report says that nothing was observed. A new source usually appears within minutes of its first probe. The current totals are on the dataset page.

Scanner tags depend on operators keeping reverse DNS and registration data accurate. Some commercial scanners share networks with abuse, so a tag tells you where an IP is hosted and nothing certain about intent. When a tag and the events disagree, go by the events.

Freshness and retention

Observations are kept for 400 days. How far back a query reads depends on the plan, and the pricing page lists the window of each. Leaderboards and the CVE table cover a recent window of 24 hours to 7 days, which each of them names. Reports for an IP, a port or a network are cached briefly, so a figure can trail the running total by a few minutes.

An IP first seen months ago keeps that first-seen date on its report.

Source code

The sensor and the enrichment pipeline are public under a source-available licence: Spip-Go and Loom. The API is described in the docs.