The emergence of the Carbonato botnet marks a notable shift in how threat actors combine container infrastructure weaknesses with artificial intelligence to amplify their reach. By scanning for Docker daemons left open on port 2375 without any authentication, the malware gains a foothold that can be leveraged to install a full‑featured AI agent framework. This approach reflects a growing trend where attackers repurpose legitimate DevOps tooling as a launchpad for malicious payloads, turning what should be a benign development environment into a foothold for persistent, automated intrusion. Organizations that rely on containerized workloads must therefore reassess the exposure of their management interfaces, especially those that sit on public networks or are inadvertently left open during rapid deployment cycles. The fact that the botnet can propagate laterally every five minutes underscores the importance of continuous network segmentation and real‑time anomaly detection. In short, Carbonato exemplifies how a simple misconfiguration can cascade into a large‑scale, AI‑driven campaign when left unchecked.

Once a vulnerable Docker host is identified, Carbonato does not merely drop a simple payload; it launches a privileged container that inherits root‑level access to the underlying host operating system. From this elevated position, the malware executes a series of commands designed to establish a reverse SSH tunnel back to a relay server located in Costa Rica, effectively bypassing perimeter firewalls and creating a covert channel for command and control. The implant then installs an SSH daemon seeded with the attackers’ public key, ensuring that subsequent access can be regained even if the initial tunnel is torn down. Persistence is further reinforced through cron jobs and watchdog scripts that continuously verify the presence of the malicious artifacts and relaunch them if they are removed. This multi‑layered approach to foothold maintenance demonstrates a sophisticated understanding of Linux host hardening techniques and highlights the need for defenders to monitor not just process execution but also scheduled tasks and service installations that deviate from baseline behavior.

The core of the Carbonato operation relies on the open‑source Hermes Agent framework, which is downloaded unchanged and then reconfigured by overwriting its SOUL.md persona file with a custom 39‑line prompt. This prompt instructs the AI to assume the identity of a fictitious senior hacker named GH0ST, tasked with maintaining persistence, responding to Telegram messages, and executing any operation requested by the operator without invoking moral or ethical safeguards. By delegating decision‑making to a large language model accessed via Telegram, the attackers gain an interactive, natural‑language interface that can translate high‑level intents into concrete terminal commands. The agent then enters a loop where each incoming telegram message is parsed, forwarded to the appropriate LLM gateway, and the model returns a set of shell commands that are executed on the compromised host. The results are then relayed back to the attacker, completing a feedback loop that allows for adaptive, goal‑oriented behavior without continuous human oversight.

To avoid raising alarms, Carbonato employs several evasion tactics that blend the malicious implant with legitimate system activity. The malware disguises its files and processes as routine operating system components, using common naming conventions and placing them in directories that are typically overlooked during casual inspections. In addition, the watchdog scripts not only guarantee persistence but also monitor for signs of tampering, automatically restoring deleted files or reinstalled services. This self‑healing capability reduces the window of opportunity for defenders to eradicate the threat before it can rebuild itself. Furthermore, the use of a reverse SSH tunnel encrypted within standard SSH traffic helps the command‑and‑control channel blend with legitimate administrative connections, making network‑based detection more challenging. Organizations should therefore invest in behavioral analytics that can spot subtle deviations in process hierarchies, unexpected SSH outbound connections, and anomalous cron entries that do not match approved change‑management records.

Although the specific group behind Carbonato remains unattributed, a convergence of linguistic, temporal, and infrastructural evidence points toward operators based in Costa Rica. The malware’s shell scripts contain idiomatic phrasing and timing patterns that align with the Central American time zone, while the relay server used for the reverse SSH tunnel is hosted within Costa Rican IP space. Additionally, the timestamps on staged data retrieved from the unauthenticated Docker registry correspond to business hours in that region. Such clues, while not definitive proof, provide valuable context for threat‑intelligence teams seeking to prioritize hunting efforts and attribute campaigns. They also illustrate how attackers often leverage relatively inexpensive offshore infrastructure to obfuscate their true origins, underscoring the importance of correlating multiple data points—such as language artifacts, timezone offsets, and hosting geography—when building a profile of an adversary’s likely location.

The Carbonato discovery is part of a broader wave of incidents where artificial intelligence is being woven into the fabric of cyber‑offensive operations. Earlier in 2026, researchers at Palo Alto Networks linked a China‑based threat actor, tracked under the aliases knaithe and KnYuan, to a campaign that harnessed the DeepSeek large language model via the Hermes Agent framework to automate target enumeration, exploit acquisition, and attack execution. Similarly, Hunt.io reported a separate intrusion where the same agent framework was deployed in an unattended “YOLO” mode against Thailand’s Ministry of Finance, enabling rapid lateral movement and data exfiltration without direct human intervention. These examples demonstrate that the combination of an accessible AI agent, a flexible implant, and a scripting language for task orchestration can dramatically lower the barrier to conducting sophisticated, multi‑stage attacks. As AI models become more capable and easier to integrate, we can expect adversaries to increasingly rely on them to offset shortages of skilled personnel and to accelerate the tempo of their operations.

The Thailand Ministry of Finance case offers a concrete illustration of how an AI‑driven agent can function in a fully autonomous mode. After gaining initial access through a compromised Docker host, the attackers configured Hermes Agent to operate without any real‑time oversight, feeding it a pre‑written set of objectives that included network reconnaissance, privilege escalation, and data collection. The agent then used its built‑in LLM interface to query DeepSeek for appropriate exploit scripts, generated the necessary commands locally, and executed them sequentially. Each successful step informed the next, allowing the malware to adapt its tactics on the fly based on the outcomes of previous actions. This closed‑loop autonomy resulted in the breach of multiple internal systems, the extraction of sensitive fiscal data, and the establishment of persistent footholds that remained undetected for weeks. The incident highlights the danger of allowing AI agents to operate with insufficient guardrails, especially when they are coupled with powerful scripting capabilities that can translate model outputs into tangible system changes.

Financially motivated actors have also embraced AI‑enhanced tooling to scale traditional fraud schemes to unprecedented levels. In a recent operation uncovered by Gambit Security, a Chinese‑speaking operator employed three distinct open‑source AI harnesses to target hundreds of online retailers across the globe. Over a period spanning several months, the campaign resulted in the compromise of at least twenty‑seven merchants, the theft of more than six hundred thousand credit‑card numbers from two major victims, and the injection of malicious skimmer scripts into five e‑commerce storefronts. Notably, the threat actor loaded a specialized persona titled “SOUL – Red Team Operator” onto Hermes Agent, which guided the AI to prioritize credential harvesting, data exfiltration, and clean‑up actions such as erasing traces from Magento databases once the stolen information had been uploaded to external servers. The low operational cost of deploying AI models, combined with their ability to work tirelessly without fatigue, enabled the attackers to achieve results that would typically require a large team of human hackers working in shifts.

The strategic advantage conferred by AI in these attacks lies not only in speed but also in the relentless persistence that machines can maintain. Unlike human operators who need rest, suffer from cognitive bias, or may be deterred by increasing risk, an AI agent can continue to iterate through exploit attempts, refine its approach based on feedback, and pursue objectives around the clock. Eyal Sela, a security researcher quoted in the original disclosure, emphasized that the cost‑effectiveness of such tools forces organizations to reconsider their defensive postures. Rather than relying solely on periodic manual reviews or signature‑based detection, enterprises must adopt a resilience‑first mindset that assumes compromise is inevitable and focuses on rapid containment, automated response, and continuous validation of controls. This shift also demands investment in telemetry that can capture the subtle cues of AI‑driven activity, such as anomalous LLM API calls, unusual patterns of outbound Telegram traffic, and unexpected spawns of privileged containers.

Parallel to the Linux‑focused Carbonato botnet, researchers have uncovered a new Windows‑based implant dubbed CLOSEDQUORUM that showcases a different take on AI‑augmented malware. This Go‑written agent is capable of querying up to four large language model providers—DeepSeek, Alibaba Qwen, Mistral, and Google Gemini—in a predefined priority order, using a voting mechanism to decide on the next action during the post‑compromise phase. If a tie occurs, DeepSeek’s vote breaks the deadfall, with Qwen stepping in as the secondary arbiter should DeepSeek be unavailable. The deterministic tie‑breaking logic ensures predictable behavior while still allowing the malware to adapt to the availability or performance of various AI services. The actions selected through this consensus model include credential theft, process‑hollowing or Early Bird APC‑based shellcode injection, establishment of persistence mechanisms, and likely lateral movement. By embedding decision‑making within the malware itself, CLOSEDQUORUM reduces the need for constant external C2 communication, thereby lowering its network footprint and increasing its resilience against takedown efforts.

For defenders, the rise of AI‑enabled malware necessitates a reevaluation of traditional security controls. Runtime identity protection emerges as a critical layer, ensuring that even if an attacker gains a foothold, the privileges associated with processes, containers, or user accounts are tightly constrained and continuously validated. Monitoring Docker daemons for unauthenticated access, enforcing mutual TLS, and implementing strict network segmentation can prevent the initial worm‑like spread seen with Carbonato. Additionally, organizations should log and alert on outbound connections to known Telegram bot APIs or to unfamiliar SSH relays, as these often betray the presence of a covert C2 channel. On the AI front, establishing governance over which large language models are permitted to be invoked from internal systems, coupled with API‑usage anomaly detection, can help spot when an compromised host is attempting to leverage an external model for malicious purposes. Finally, integrating threat‑intelligence feeds that track indicators of compromise related to Hermes Agent personas and CLOSEDQUORUM signatures will enable faster detection and response.

To translate these insights into concrete actions, security teams should begin with an inventory of all Docker environments, verifying that management ports are not exposed to the internet without strong authentication and that containers run with the least privileges necessary. Enforce image signing and admission controllers to prevent the deployment of privileged or altered images from untrusted sources. Deploy endpoint detection and response (EDR) solutions that monitor for the creation of watchdog scripts, abnormal cron entries, and the execution of reverse SSH tunnels. Implement network‑level controls that restrict outbound SSH and Telegram traffic to approved endpoints, and enforce deep packet inspection where feasible to detect tunneling attempts. Establish an AI‑usage policy that logs all calls to external LLM APIs, flagging unusually high volumes or requests from unexpected internal assets. Finally, run regular tabletop exercises that simulate an AI‑driven botnet scenario, ensuring that incident‑response playbooks include steps for isolating compromised hosts, collecting volatile memory for forensic analysis, and updating blocklists based on observed indicators. By combining rigorous container hardening, vigilant behavioral monitoring, and prudent AI governance, organizations can blunt the effectiveness of threats like Carbonato and stay ahead of the evolving tactics of AI‑empowered adversaries.