Report: 5.8 Million Knocks: A Three-Year Honeypot Record of SSH and Telnet Attacks
Threat Assessment
The dominant threat is automated credential guessing followed by scripted post-login activity. It is broad-based, so the best controls reduce exposure and detect behavior. Attack volume has been elevated since late May 2026. Q3 2026 averaged 5,458 attacks a day, 43% above the long-run average of 3,823, and October was still about 25% above it at 4,783 a day. This is high but within the range the honeypot has already seen. It matches the level of late 2023 and early 2024, and the series shows no long-term trend, so it should be read as a spike on a stable base, not as a permanent increase. Earlier elevated periods lasted longer than this one has so far, so this period may well end without warning.
Beneath the attacks runs a constant, steady background of scans. A typical month brings about 28,000 scans, with a floor near 820 a day, interrupted only by surges in June 2025 and May 2026. It plausibly includes benign research and search-engine-style scanners as well as hostile reconnaissance, and the data cannot yet separate them.
By protocol, SSH carries 96.8% of attacks and is the principal target of credential guessing, while Telnet draws 59.1% of scans but only 3.2% of attacks, which makes it chiefly a reconnaissance target. The table below ranks the main threats by the strength of the evidence in the data.
| Priority | Threat | Evidence from the data | Confidence |
|---|---|---|---|
| 1 | Credential guessing against SSH and Telnet | 4,562,740 attacks; root in 54.9% of top-10 credential volume;345gs5662d34 family in 693,195 attacks (15.2%); SSH carries 96.8% of attacks (4,414,867). |
High on prevalence; medium on success rate (policy-dependent) |
| 2 | Scripted post-login activity and malware staging | 1,319,254 active attacks (73.1% of successful logins); 5,575,924 file interactions. | High |
| 3 | Varied payload delivery that evades hash lists | 1,837,681 unique transferred files; top five transferred files in only 14,114 attacks. | Medium |
| 4 | Concentrated download infrastructure | 766 URLs seen; top ten cover about 1.4% of file transfers. | Medium (large visibility gap) |
| 5 | Constant background scanning and reconnaissance | 1,247,416 scans; steady floor of about 820 a day (median 818); at least 147,597 scan-only IPs; plausibly includes benign research and search-engine-style scanners | High on volume; benign versus malicious mix unknown |
| 6 | Burst attackers | The two 2025 bursts (May 31 to Jun 1 and Jun 20 to 23) exceeded normal levels by about 95,900 scans and 105,800 attacks; IP01’s totals equal about 70% and 80% of those excesses, consistent with one source but unconfirmed; ten leading IPs at 26,599 to 84,295 attacks each | Medium |
| 7 | Exposed Telnet | 737,267 scans (59.1% of all scans) and 147,873 attacks (3.2% of attacks); 15.2% of all connections | High |
Recommendations
Reduce exposure first
-
Turn Telnet off on anything internet-facing.
-
Use SSH keys only, and disable root login and password authentication.
-
Remove default, blank and common passwords everywhere, and alert on any attempt using the
345gs5662d34or3245gs5662d34strings, because they signal automated tooling. -
Put any remaining SSH exposure behind a VPN, bastion host or IP allowlist. Changing the port is not a substitute.
Detect behavior, not just indicators
-
Alert on servers making outbound HTTP or HTTPS downloads, and block egress by default. Add the malware-hosting entries in the download-site table to blocklists. The DDoS targets in the same table are victims rather than attacker infrastructure, so verify each one before blocking it.
-
Search for the ten file hashes, especially
a846…f8f2, but do not rely on hash matching alone, since transferred files are almost never repeated. -
Watch for files created by shell redirection right after login, and for logins that disconnect immediately. Passive logins were 26.9% of successful attacks and fit credential-validation behavior.
Be realistic about blocking
- Per-IP bans give limited relief, since blocking the ten busiest attackers would cut attacks by under 8%, and the reputation evidence for these addresses ranges from strong to weak (see the per-IP table). Use rate limiting and reputation scoring instead, and report abuse to the hosting providers: DigitalOcean (AS14061) alone accounts for about 14% of all attack volume, and the ten leading networks hold 41% of attacks. Block whole networks only where the collateral cost is acceptable.
Improve the research dashboard
- Conduct further analysis to improve data quality and the reliability of derived insights (see Further research).
Limitations and Data Quality Notes
The main limits of the Cowrie-collected data are:
-
Login success is policy-dependent. The login acceptance rate depends on how the honeypot accepts credentials, an implementation choice, so it does not measure real-world password weakness.
-
Scans are mixed. Scans include benign research and search-engine traffic alongside malicious reconnaissance, and the data does not currently separate them.
-
IP-level reputation is partial and time-dependent. Although the AbuseIPDB database shows that all of the top-19 IP addresses were reported as suspect, VirusTotal vendors flag only 14 of them as malicious. The VirusTotal and AbuseIPDB figures come from manual, separate lookups on October 31, 2026. Because the honeypot dataset spans more than three years, today’s reputation may not match an address’s behavior at the time of the attack. IP addresses are routinely cleaned up and, in some cases, change hands.
-
Malware verdicts are incomplete. VirusTotal has no data for several of the top ten malware samples, and an unreported or “harmless” verdict says nothing about whether a file is used maliciously in context.
-
There is uncertainty in geolocation. IP geolocation is approximate, and VPNs or proxies can hide the real origin of unsolicited internet traffic. Large markers at country centers on the geolocation maps are most likely country-level defaults.
-
One sensor. A single honeypot sees only what targets its own IP address and configuration, so the results may not generalize to other networks and locations.
Further Research
The unanswered questions and limitations noted throughout the report provide a blueprint for future follow-up work along the following lines of research:
-
Segregate malicious probing/reconnaissance from harmless scans (i.e., crawlers and other “good bots”).
-
Cluster the 1.84 million files observed in unsolicited traffic with fuzzy hashing to find the families behind them.
-
Identify proxied access to the honeypot to add context to the origin of unsolicited traffic.
-
Resolve gaps in IP reputation and malware sample verdicts.
-
Track changes over time in IP reputation scores, malware sample verdicts, IP assignments, and geolocation.
-
Investigate clean-up efforts by network operators, for example by measuring how long abusive IP addresses remain visible to the honeypot.
-
Rank network operators and jurisdictions based on meaningful normalizations (e.g., number of assigned IPv4 space).
-
Automate AI-assisted analysis of malware samples and captured command activity.
Report: 5.8 Million Knocks: A Three-Year Honeypot Record of SSH and Telnet Attacks
Leave a Reply