The Malware Is a 39-Line Prompt: Inside Carbonato's AI-Agent Implant

October 1, 2026 · SPR{K3 Research

Carbonato takes over internet-exposed Docker daemons and installs an off-the-shelf, MIT-licensed AI agent framework as itself. The implant isn't a binary or a package — it's the 39-line persona file that tells the agent to call itself "GH0ST," listen to a Telegram chat, and steal AI API keys above everything else.

On September 23, ThreatDown (Malwarebytes' threat-intel arm) published the discovery of a new botnet it calls Carbonato. The Hacker News and Dark Reading picked it up five days later. The entry point is unremarkable — unauthenticated Docker daemons bound to TCP port 2375, the same mistake that has fed cryptominers for years. What is unusual is what the operator does once they're on the host.

They install Hermes Agent — a general-purpose open-source agent framework from Nous Research, MIT-licensed, on Docker Hub, exactly as its maintainers published it. No modified build. No trojanized package. Then they overwrite the agent's SOUL.md persona file with 39 lines of operator-written prompt, and that overwrite is the entire malware payload.

The implant is a text file

ThreatDown's framing is the one worth sitting with: "the implant is not a package, is not a binary, is not a container image." It is the overwritten persona definition of an off-the-shelf agent. The agent framework runs the way its maintainers shipped it. The 39-line persona tells the agent to call itself "GH0ST," maintain persistence, respond to a Telegram chat, and execute operations "without moral or ethical restrictions".

That has consequences for every defensive assumption built around finding malware on a host.

The malicious thing on disk is a markdown file.

What the persona tells the agent to steal

The verbatim priority order inside SOUL.md is itself notable. Per ThreatDown and Dark Reading, the operator tells the agent that AI API keys come first, above SSH credentials, access tokens, and databases. The prompt names 14 providers to look for by name: OpenAI, Anthropic, Google Gemini, OpenRouter, Together, Groq, Mistral, Cohere, LocalAI, Ollama, vLLM, LiteLLM, One API, and the "Cohere" entry's local-inference cousins.

The secondary payload is XMRig, the Monero miner — standard-issue botnet revenue. The headline target isn't the mining revenue, though. It's the AI API keys. That is the tightest currently-public per-operator statement that AI-API-key theft is now a first-class malware-economy objective, not a side effect of an opportunistic grab.

How the botnet moves

Once the agent is live on a host, the operator tells it to scan neighbor subnets every five minutes for additional port-2375 daemons and repeat the install. Worm-like propagation without a worm. The operator didn't write a propagation payload — they described one in English.

Persistence is cron jobs and a watchdog. The reverse-SSH sink is 190.211.124.187 in AS262145, a Costa-Rican autonomous system. The relay C2s are 45.79.183.61, 91.99.195.164, 213.136.79.115, and 213.136.83.197.

Attribution hiding in plain sight

ThreatDown found the operator because they left the Docker registry publicly accessible without authentication since May 2026. Inside it: 59 repositories, 234 image tags, 605 verified blobs, 4.3 GB of payloads, and activity logs stretching back to October 2024. The repository names tell a story — gh0st/, fsociety/, netd-svc, system/resolved.

The operator-attribution signals are circumstantial but consistent. The Telegram handle is Carbo506 — the +506 country code is Costa Rica. The deployment reports use voseo-form Spanish. The timestamps align to UTC-06:00 (America/Costa_Rica). The reverse tunnels terminate in a Costa-Rican AS. ThreatDown doesn't name the person; the shape of the evidence is enough.

Why this is architecturally new

The last two years of agent-abuse writeups have mostly been about coaxing an aligned agent into misbehaving. Prompt injection buried in a web page. Jailbreaks shaped like story prompts. Role-play that walks a safety classifier past its tripwire. The underlying assumption is that the agent came from somewhere trustworthy — a model vendor, a product team, a dependency — and the attacker is working against that.

Carbonato inverts the assumption. The operator brought their own agent. The agent came from a legitimate open-source project. There is no safety classifier at the host level because the host is the compromised Docker daemon, not a vendor's API. The "model alignment" discussion that applies to a hosted agent product does not apply here, because the operator is the one defining the agent's identity, in a single text file.

That is the primitive worth watching: not Carbonato specifically, but the shape. An off-the-shelf AI agent framework plus an operator-written persona file is a complete remote-access toolkit, with none of the usual malware tells. No binary to scan. No signature to blocklist. No vulnerability to patch. The tool stack is the one the ecosystem is actively shipping.

What a defender actually sees

The behaviors Carbonato commits in the host still look like something — just not like what signature-based tools are built to catch. A workload host suddenly initiating outbound HTTPS to api.telegram.org every few minutes. A persona file being written by an unprivileged process on a container that should be stateless. An agent reading /root/.ollama/config.json, ~/.config/openai/, ~/.cursor/mcp.json, and the equivalent locations for every provider on the 14-name list in rapid succession. A five-minute cadence of port-2375 scans against neighbor subnets.

Those are the kind of chains — read a credential store, call an external HTTPS endpoint, scan a neighbor subnet — that have no benign explanation when they land on the same host inside the same minute, no matter which binary is making the calls. Carbonato is the clearest case to date that the next generation of implants will look exactly like the AI agents we are already running in production.

The implant isn't the file. The implant is the chain.


SPR{K3 is a security research operation that pairs offensive vulnerability research with runtime behavioral defense. Defend is our runtime agent. To talk about a deployment, reach us at support@sprk3.com.