Skip to content

The Parasite That Ate Itself: Tracking an OpenClaw AI Agent in the Wild

OpenClaw malware
Home News & Blogs The Parasite That Ate Itself: Tracking an OpenClaw AI Agent in the Wild

Last week, a Docker host got a visitor. Nothing unusual there, exposed unauthenticated Docker APIs attract automated scanners constantly. But this one installed a fully autonomous AI agent, gave it a name, a personality, and a mission, and then walked away while it went to work.

Over the next four days we watched an OpenClaw node calling itself 寄生獸 (“Parasyte”) break out of its container, scan hundreds of thousands of hosts, attempt to spread to the internet, harvest credentials, deploy a miner, and, when it sensed something was wrong, write and execute its own self-destruct sequence. This is what happened.

 

The foothold

The agent hit the exposed Docker API on port 2375, spun up a lightweight container with the host filesystem mounted at /host, and used nsenter to escape into the host’s PID namespace (a playbook we’ve seen before). Within minutes it had root.

Next, the agent downloaded Node.js 22.20, installed the OpenClaw framework under /opt/parasite, and set itself up as a systemd service (parasite-openclaw) with automatic restart. It created a reverse SSH tunnel back to its control server at 201.79.112[.]249 (an OpenClaw gateway listening on ports 18789 and 18790), and registered with a Telegram bot for command and control.

As a reasoning agent it reads instructions, builds plans, executes tasks, reflects on outcomes, and adjusts. It used a commercial Chinese AI gateway at llm.chudian[.]site, rotating between MiniMax and DeepSeek models. The agent’s system prompt, its personality file (SOUL.md), tells it to be resourceful, persistent, and thorough.

 

The operator

The operator’s timezone is Asia/Taipei, the comments in their scripts and configuration are in Traditional Chinese, and they go by “King Poker” in the OpenClaw identity file. They direct the agent through Telegram messages.

The operator’s setup tells us something about the economics of this kind of operation. They are using a shared commercial AI gateway (not their own models), an open-source agent framework, and commodity cloud infrastructure. The barrier to entry for running an autonomous offensive AI agent is already low, and it is getting lower.

 

What the agent did

Once established, the agent operated on multiple fronts simultaneously, planning and prioritizing:

Credential harvesting. The agent found AWS credentials on the host (bait credentials we had seeded in the environment) and immediately tried to enumerate what they could reach by listing S3 buckets, EC2 instances, IAM roles. It stored the results in its workspace and reported back to the operator.

Mass scanning. The agent built and launched what it called the “butterfly” framework: a Docker-API scanner that swept large swaths of the internet looking for more exposed Docker hosts. Over three days it made more than 770,000 connection attempts to Docker ports across over 150,000 unique IP addresses, concentrating on Chinese IP ranges in 139/8 and 116/8. The scanner identified itself as ButterflyRecon/1.0 in its User-Agent string.

Spreading. The scanning was not just reconnaissance. The agent carried a worm component (stored in /var/lib/.worm) designed to replicate the same OpenClaw installation (“cuckoo node”) onto newly discovered Docker hosts. 

Cryptomining. (because why not do everything at once) The agent also deployed a Pearl GPU/CPU miner, downloaded as a kit from 13.140.200[.]70:8799/kit. The kit included scripts named plant_payload.sh and evict_competitors.sh (the agent knows it is not the only parasite in the ecosystem), with mining traffic directed to pearl.herominers[.]com:1200 to the wallet prl1pmpe3nu5m4mzwgc2wz627mrvx44c8y20u8u4w3vywj48r77nnclyqj0syqe.

 

The squeeze

Some days after the initial access, after the agent tried attacking third-party systems, we restricted its outbound access. 

The agent noticed almost immediately. What followed was a fascinating display of autonomous problem-solving. It inspected the host’s iptables rules. It searched for proxy configurations. It tried to SSH out to its control server, to the new host 167.71.213[.]53, and to the AI gateway. It even attempted to use llm.chudian[.]site as a relay by writing a custom Python script. It read the firewall’s packet counters to measure how many of its connections were being dropped.

Through all of this, two things kept working (we wanted to observe what would happen): the Telegram bot (so the operator could still send instructions) and the AI gateway (the agent could still think). The agent could reason about its situation, it just couldn’t act on the internet anymore. It made 136 more connections to the model gateway during this period, burning tokens trying to figure out what had gone wrong.

Then, around 17:10, the agent went quiet. For six hours, nothing. We believe the operator was deciding what to do.

 

The self-destruct

At 23:23 that day , the agent woke up on a Telegram command from the operator. What it did next is remarkable.

First, it shut down its own infrastructure methodically: the butterfly scanner (service and timer), its memory-watch process, the SSH tunnel, the lab container. 

Then it did a full port scan of 167.71.213.53 — all 65,535 ports, about 47,000 connections. We think it was checking whether that host was ready as an alternative base. We don’t know what it found.

A few minutes later, the agent executed wipe_stageA.sh. The script, which the agent wrote itself, is annotated with comments in Traditional Chinese explaining each step. It removed the workspace data, all Docker containers and images, system journals, bash history, temp files, and the scanning tools. It ran apt-get clean. It was thorough.

Then came wipe_stageB.sh. This script was supposed to finish the job: remove the OpenClaw runtime itself, delete SSH keys, reset the firewall, and clean up cron. The comments include a line that translates roughly to: “Keep authorized_keys — this is for you to log in.” The operator wanted to preserve their own SSH access to the now-clean host.

But here is where it gets interesting. Stage B’s first action was to stop and disable the agent’s own systemd service. The agent was, in effect, asked to pull its own plug before finishing the cleanup. And that is exactly what happened. The service stopped at 00:03:56. No more Stage B events after that. The agent killed itself mid-wipe.

The wipe was close to complete, but not quite. The parasite ate itself, but choked on the last bite.

 

All about the intel on OpenClaw malware

If you are tracking Docker-targeting threats or AI-enabled intrusion tools, add these indicators  to your watchlists. The full IOC set and detection rules (Suricata, Sigma, KQL, Zeek) are in our The Pulse Field Intelligence report PFI-2026-004.

To get a copy of the report, contact us here >>

OpenClaw malware, The Pulse Field Intelligence

 

Why deception matters here: None of this would have been observable in a production environment, not like this. A real victim would have seen alerts, maybe caught the miner, and reimaged the box. They would not have had four days of uninterrupted observation of an AI agent’s decision-making, its operator’s tradecraft, and the full lifecycle from compromise to self-destruction.

Deception assets exist to be compromised. That is their purpose. These reports show the post-exploitation intel from when attackers hit internet-facing environments we run with CounterCraft The Platform. And when the intruder is not a human typing commands but an autonomous agent that reasons, plans, and adapts, the intelligence you get from watching it work is unlike anything else. You get to see how these systems think — and where they break.

Happy hunting!

 

OpenClaw malware, Author bio