Executive Summary
In late 2026, researchers identified CARBONATO, a sophisticated Docker-based botnet that utilizes the open-source Hermes Agent AI framework to automate malicious activities. The campaign, linked to operators likely based in Costa Rica, targets unauthenticated Docker daemons on port 2375 to establish a foothold. Once inside, the botnet uses privileged containers to gain host-level access, maintains persistence via multiple Linux system hooks, and propagates autonomously to adjacent networks.
The technical innovation of CARBONATO lies in its use of an AI agent persona named 'GH0ST.' By overwriting the agent's instructions, attackers leverage Large Language Models (LLMs) to interpret terminal output and execute complex post-exploitation tasks directed via Telegram. Notably, the threat actors prioritize the theft of AI API keys from providers like OpenAI and Anthropic, reflecting a shifting priority in modern credential theft aimed at fueling further AI-driven attacks.
This activity represents a growing trend of 'autonomous' cyberattacks where human intervention is minimized. Affected organizations face risks of persistent infrastructure compromise, lateral movement, and the exfiltration of high-value secrets. The ability of the botnet to scan and infect new hosts every five minutes emphasizes the need for immediate remediation of exposed container APIs.
Key Details
Threat Name
CARBONATO AI Botnet
Affects
—
Adversary
CARBONATO Other Adversaries and Aliases: Akira; knaithe
MITRE Techniques
Malware/Tools
CARBONATO, Hermes Agent, XMRig, Strix, Cairn, CLOSEDQUORUM
