The Cowrie SSH honeypot, deployed alongside a portfolio server, recorded 24 218 sessions from 683 addresses in its first five days. The practical takeaway here isn’t the record number of attempts: one bot family generated nearly 88% of the sessions, so connection counts alone can easily exaggerate the diversity of attacks.
The person who built the setup deliberately made it a fake Linux system: it accepts almost any password and records visitors’ actions, but there’s no real OS behind the SSH session. This format makes it possible to observe the behavior of scanners and bots without giving them a working environment. And yes, when nearly nine out of ten records belong to one automated guest, the stats need more than a counter—they need a bit of sorting, too.
The honeypot is isolated, and its logs are scrubbed
Cowrie runs in an unprivileged container, in read-only mode and with outbound connections blocked. So even if a visitor enters a command that would normally contact an external server, the honeypot itself can’t make that connection. Other decoys are deployed alongside it—for Redis, MySQL, and Elasticsearch—and CrowdSec filters the site’s incoming traffic.
A separate part of the setup is the so-called tar pit, which mimics an SSH port but delays the client before the usual banner. The honeypot’s logs are aggregated every five minutes. During processing, commands are scrubbed: links, IP addresses, and random filenames are replaced with markers such as <url> і <file>, and attackers’ addresses are not published.
This setup shows not only how many connections come in, but also what happens after they connect. The interval between new sessions is usually about five seconds, although the server, according to its author, wasn’t advertised anywhere in particular. It’s found by automated scanners sweeping through internet addresses.
Passwords are simple, but sessions aren’t independent
Among the most frequently entered passwords are 123456, 123, 1234 and 12345678Bots try the logins root, ubuntu, admin, deploy і postgres. The log also included claude; автор припускає, що автоматизовані сканери шукають облікові записи, пов’язані з інструментами розробки на основі ШІ. Сам факт появи такого логіна не доводить, що його перевіряє окрема спеціалізована кампанія.
The biggest contribution to the overall statistics came from a family associated with 18 addresses: it generated 21 217 sessions, or 88% of the recorded total. After logging in, the bot ran one command— uname -s -v -n -r -m —and disconnected. It requests system information, including the system type and kernel version.
The author interprets this as possible inventory-taking: the bot checks whether the password worked and what kind of system responded. This behavior could fit with compiling a list of accessible servers to revisit later or pass on to others. It’s a reasonable theory, not an established operator’s intent. The concrete takeaway for analysis is this: 21 thousand sessions does not mean 21 thousand different attackers or tactics.
Another bot replaces SSH keys
By number of addresses, another family stood out—it covered 291 addresses and acted much more aggressively. The script removed protection from the directory .ssh, deleted it, recreated it, and wrote its own key to authorized_keys. In the honeypot, this scenario is only recorded; on a real machine, this sequence could delete existing keys and leave access through the attacker’s key.
cd ~ && rm -rf .ssh && mkdir .ssh && echo <ключ> >> .ssh/authorized_keys
This has an important implication for responding to a possible compromise: simply changing the password isn’t enough if a key has already been added to the account. The author recommends checking authorized_keys, cron jobs, and the user list. Deleting the entire directory .ssh could also lock out the server’s owner and other users, so that command on a real host is not cosmetic cleanup.
Because outbound connections are blocked, the honeypot doesn’t download files when bots issue commands to do so. At the same time, visitors themselves sent 22 files there, 17 of which were Linux programs for different architectures, including x86-64, ARM, and AArch64. Several files showed signs of XMRig, a cryptocurrency miner. They were analyzed as data and not run, so the conclusion about mining is based on the signs found, not on executing the programs in the honeypot.
The tar pit delays connections, but doesn’t prove service availability
The delay mechanism works in accordance with RFC 4253: before the banner appears, the server can send lines, and the client has to wait. The pit sends a random line every ten seconds and holds a connection for no longer than a day. One week after launch, according to the author’s counter, it had accumulated over 2 000 hours of delay and more than 3 400 connections. An individual connection lasted at most 24 hours, and usually a few dozen bots remained in the pit at once.
The author estimates that if each of these bots tried a password once per second, the delay would amount to about 7,5 million attempts that never happened. That’s a hypothetical conversion, not a measured number of thwarted attempts: the mechanism simply keeps the connection open, and this estimate doesn’t establish how bots would actually behave in such a scenario.
The more surprising practical problem wasn’t an attack, but the setup’s own configuration. For two days, the counters for the tar pit and decoys stayed at zero, even though the services were running, the firewall rule was present, and its counter was increasing. A local check passed, too.
A check via tcpdump showed that requests were getting through, but responses weren’t coming back. The cause turned out to be a second firewall: allowing traffic in one rule didn’t override the block in another. The first-level counter honestly showed its section of the route, but not the packet’s entire journey—firewalls, it turns out, don’t exchange authority by default.
So a successful test from the server itself doesn’t confirm that the service is accessible from outside. Likewise, the counter for one rule doesn’t prove that traffic got all the way to the end client. In this case, the answer came from an external connection together with packet observation. The author also plans to check whether bots change their behavior when the honeypot changes its server name and system version; the description doesn’t yet include results from that experiment.









