Hugging Face's logo

Datasets:
zer0contextlost
/
flypaper

Modalities:
Text
Size:
10K - 100K
Tags:
cybersecurity
honeypot
threat-intelligence
cowrie
ssh
telnet
License:
Dataset card Data Studio Files
xet
 Community
flypaper / README.md
zer0contextlost's picture
zer0contextlost
card
e8e07d5 verified
preview code
|
Raw History Blame Contribute Delete
5.87 kB
metadata 
license: cc-by-4.0
pretty_name: 'Flypaper: open honeypot data'
size_categories:
  - 100K<n<1M
tags:
  - cybersecurity
  - honeypot
  - threat-intelligence
  - cowrie
  - ssh
  - telnet
  - malware
  - intrusion-detection
configs:
  - config_name: sessions
    default: true
    data_files: sessions/*/*.jsonl.gz
  - config_name: web
    data_files: web/*/*.jsonl.gz
  - config_name: campaigns
    data_files: campaigns/campaigns.csv
  - config_name: malware
    data_files: malware/samples.csv

Flypaper: open honeypot data

Everything that attacked three internet-facing honeypots on one residential IPv4 address in the United States: SSH and Telnet sessions with every login attempt and command, malware hashes, web exploit requests, and the attack campaigns they group into. It's rebuilt nightly from the raw logs and released under CC-BY-4.0.

Current contents (2026-08-10 to 2026-11-12, UTC):

SSH/Telnet sessions 324,805 (250,780 honeyfleet, 74,025 Cowrie)
login attempts 310,665
commands typed 467,411
files dropped (hashes) 14,120
web requests 70,703
IPs that hit the Cowrie and web sensors (since 2026-08-09) 4,113
campaigns 2,981
malware samples triaged 239

Sensors

sensor what it is data from
honeyfleet SSH honeypot: a fork of Cowrie with a lived-in fake filesystem, rewritten common commands, and a per-session classification of the client (bot or human, typing cadence, command skeleton) 2026-08-10
cowrie stock Cowrie on SSH and Telnet 2026-10-17
web low-interaction HTTP catch-all with a fake WordPress login form 2026-10-19

Raw Cowrie logs before 2026-10-17 and web logs before 2026-10-19 were not kept. For those weeks attackers.csv.gz still has per-IP totals, usernames and passwords going back to 2026-08-09.

Every sensor accepts some logins on purpose, so success: true means the honeypot let the attacker in. It doesn't mean the password is real anywhere.

Files 

sessions/YYYY-MM/sessions-YYYY-MM-DD.jsonl.gz   one JSON object per SSH/Telnet session, by UTC start day
web/YYYY-MM/web-YYYY-MM-DD.jsonl.gz             one JSON object per HTTP request
attackers.csv.gz                                per-IP history for the Cowrie and web sensors since 2026-08-09
campaigns/campaigns.csv                         campaigns found by grouping sessions on behaviour
campaigns/campaign_ips.csv.gz                   which IPs took part in which campaign
campaigns/campaign_daily.csv                    sessions and new IPs per campaign per day
malware/samples.csv                             static triage of captured files (hashes only)
malware/iocs.csv                                C2 addresses from those samples, confirmed by decompiling or from the dropper script
stats.json                                      the numbers above

Only finished UTC days are published. A day's file can still change on a later run if a late log rotation adds sessions to it.

Session fields

field
session, sensor, protocol session id (unique per sensor), cowrie/honeyfleet, ssh/telnet
start, end, duration_s ISO-8601 UTC; end is null if the close event was lost
src_ip, src_port the attacking address
asn, as_name, as_country from iptoasn.com at export time, not at the time of the attack
known_scanner label (e.g. Censys, Shodan) when reverse DNS identifies a research scanner
client_version, hassh, ssh_key_fingerprints, ja4 client fingerprints
logins[] t, username, password, success
commands[] t, input, ok (false if the honeypot didn't implement some part of the line, true if it ran all of it, null if it logged neither)
downloads[], uploads[] sha256 and the path or filename the attacker used
failed_downloads[] URLs the honeypot could not fetch
tunnels[] direct-tcpip forwarding requests: the host and port the attacker tried to reach through us
classification honeyfleet only: actor, actor_confidence, cadence, cadence_confidence, command_skeleton

Web records have ts, src_ip, the ASN fields, known_scanner, method, path, query, user_agent, and the username/password sent to the fake login form.

What was changed or left out

  • The sensors' own public IP is replaced with <sensor-ip> wherever an attacker typed it (it turns up as a password and in client version strings). Any private IPv4 address is replaced with <private-ip>, and domain names belonging to the operator with <own-domain>.
  • Sessions from private or local addresses (our own testing) are dropped.
  • Not published: malware binaries (the hashes are on MalwareBazaar), terminal recordings, what the honeypots printed back, payload bytes sent through tunnels, and IDS logs.

Using it 

import pandas as pd
s = pd.read_json("sessions/2026-10/sessions-2026-10-31.jsonl.gz", lines=True)
s.explode("logins").dropna(subset=["logins"])["logins"].apply(pd.Series)["password"].value_counts().head(20)

Most source addresses are compromised machines or rented servers, not the people behind them. Don't treat an IP in this data as proof of who did something.

Detection rules written from this data are in honeypot-detections. Live views of the same sensors: watch, fleet.

License and citation

Data: CC-BY-4.0. ASN data: iptoasn.com (PDDL).

zer0contextlost. Flypaper: open honeypot data. https://github.com/zer0contextlost/flypaper

Questions and corrections: open an issue.

Mirror of https://github.com/zer0contextlost/flypaper (rebuilt nightly).