Carbonato Malware Turns AI Agent Frameworks Into Attack Infrastructure
What happened
Researchers at Malwarebytes ThreatDown published findings on a botnet malware campaign called New Carbonato malware uses AI agents to hijack exposed Docker hosts that targets Docker daemon APIs left exposed without authentication. Once inside, Carbonato installs the open-source Hermes Agent AI framework and configures it with a persona named GH0ST. The agent then connects to a Telegram channel operated by the attacker, receives natural-language task instructions, translates them into terminal commands autonomously, executes those commands on the host, and reports results back. A key objective of the campaign is harvesting AI API keys and other stored credentials, giving attackers access to downstream AI services. The campaign represents a significant shift: attackers are no longer just breaching environments to steal data or deploy ransomware, but are using legitimate AI agent tooling as an autonomous command layer inside compromised infrastructure.
Why it matters
- ·AI API keys harvested by Carbonato can be used for unauthorized model access, data exfiltration through AI services, or cost fraud. Organizations must treat stored AI API credentials with the same controls applied to privileged access credentials.
- ·The use of a legitimate AI agent framework (Hermes Agent) as post-compromise tooling means existing endpoint and container detection tools may not flag the agent as malicious. Unauthorized AI agent frameworks deployed inside the perimeter now require explicit detection and classification controls.
- ·Incident response playbooks built around conventional malware rarely address AI agent abuse as a threat vector. The Five Eyes Guidance on the Careful Adoption of Agentic AI Services highlights that agentic systems require distinct containment and response procedures, and Carbonato is the first named campaign to confirm that gap is being exploited in the wild.
Governance controls affected
What to do now
- ☐Audit all Docker daemon APIs across cloud and on-premises environments and confirm none are exposed to the internet without authentication or network-level access controls.
- ☐Inventory AI API keys stored in developer environments, CI/CD pipelines, and container images. Rotate any keys that cannot be confirmed as scoped to least-privilege access, and move all keys to a secrets manager with access logging.
- ☐Add detection rules to container and endpoint monitoring for unauthorized installation of AI agent frameworks such as Hermes Agent, and flag any Telegram-based outbound command-and-control traffic originating from container workloads.
- ☐Update your AI incident response playbook to include scenarios where an AI agent framework is the execution layer for an attacker. Define containment steps specific to agentic workloads, including session termination, agent identity revocation, and downstream API key invalidation.
- ☐Review your AI system inventory for any self-hosted or open-source agent frameworks running in production. Confirm each has an authorized owner, a documented permission boundary, and an approved deployment record before the next quarterly governance review.
What to watch next
As Carbonato demonstrates a repeatable method for weaponizing AI agent frameworks, security teams and regulators are likely to pay closer attention to container security baselines for AI workloads. Compliance teams should monitor whether guidance bodies such as CISA or the Five Eyes network issue follow-on advisories specific to AI agent framework abuse. The growing body of agentic AI threat intelligence, including the Check Point 2026 Report Maps Agentic Attack Paths Compliance Teams Must Govern and similar research, suggests that container-hosted AI agents will feature prominently in future regulatory expectations around secure deployment. Organizations should also track whether AI API credential theft begins appearing in mandatory breach notification filings, which could trigger reporting obligations under existing frameworks.
Stay ahead of stories like this
Get every Global AI governance development like this one, plus the rest of the week's developments. Every Thursday.