Cyberattacks

Carbonato: an AI-agent botnet targets exposed Docker APIs

ThreatDown documents Carbonato, an implant abusing unauthenticated Docker daemons, installing Hermes Agent and seeking credentials. How to check exposure and respond.

Conceptual illustration: Carbonato: an AI-agent botnet targets exposed Docker APIs
Conceptual illustration · SecurCheck

What is confirmed

ThreatDown's 22 September research began with an unauthenticated Docker registry: 59 repositories, 234 tags and 4.3 GB of read-only evidence. The material spans October 2024 to August 2026; a 24 September news report brought wider attention.

The implant targets unauthenticated Docker APIs on port 2375. It instructs the daemon to launch a privileged container with host access, then establishes persistence and remote access.

Carbonato deploys Hermes Agent with a modified GH0ST persona. Its instructions prioritise AI API keys, then SSH credentials and tokens; scripts seek further Docker daemons on adjacent networks.

What it means

The documented entry point is an exposed Docker API without access control. The AI agent handles post-compromise tasks; the research does not establish autonomous discovery of every victim.

Host compromise can expose mounted container secrets and deployment credentials. The published evidence does not establish a global victim count.

What to do

Inventory Docker daemons and ensure port 2375 is not reachable from the internet or untrusted networks. Remove unnecessary exposure and require registry authentication.

For a suspected host, isolate it and inspect privileged containers, host mounts, SSH services, persistence and Telegram traffic; rotate affected API keys and SSH credentials after investigation.

Check a suspicious sign

External checks may reveal some signals, but investigating Carbonato requires Docker host inventory and analysis.

Sources

  1. ThreatDown — enquête originale, 22 septembre 2026