The technological significance of the Armory Castle campaign lies in its use of custom tooling to weaponize the zero-day. Google stated that the group utilized a utility dubbed BONESAW, written in Golang, to exfiltrate arbitrary Lightning Memory-Mapped Database (LMDB) files, hijack the SAML Security Token Service (STS), and create a Single Sign-On (SSO) Administrator account. The vulnerability, assigned a CVSS score of 9.8, was not patched until November 19, 2026, leaving a window of exposure during which the attackers operated. The group’s knowledge of the flaw appears to have preceded its public disclosure of a related vulnerability, CVE-2026-59309, on July 29, 2026, as the recovered BONESAW binary contained a build timestamp of July 13, 2026. While metadata can be modified, this timing suggests advanced reconnaissance and preparation.
The attack vector was methodical. After gaining initial foothold, the threat actor moved laterally to an internal Linux appliance, escalated privileges to root, and deployed a bespoke proxy tunneler called BARECLAW. This movement was part of a broader campaign that also exploited a zero-day in BeyondTrust Remote Support (CVE-2026-102764, CVSS 9.3) for reconnaissance and persistence. To locate their target, the group leveraged nmap to scan the internal environment for systems running open ports specific to VMware, notably ports 9080, 5480, and 2012. Port 2012, used as the control interface RPC for vCenter Single Sign-On (SSO), was integral to the remote exploitation of CVE-2026-123890 via BONESAW.

Once inside the virtualized environment, the attackers executed a series of actions to consolidate control. After executing a component called BEARBONES to create a listener on TCP port 33992, Armory Castle moved laterally to the vCenter appliance as the Administrator. They created a local administrator account named [email protected] and enabled Secure Shell (SSH) on vCenter via the vCenter Server Appliance Management Interface (VAMI) API through a PUT request to /rest/appliance/access/ssh. This allowed them to connect via SSH using the newly created credentials. In the final stage, the adversary enumerated the ESXi file system, mounted Virtual Machine File System (VMFS) volumes, and added an SSH public key to the root account’s authorized keys, facilitating further data theft and environmental mapping.
Concurrently, a different but related evolution in cyber operations was observed by Amazon Threat Intelligence, involving the use of an autonomous AI agent platform called Cairn. Between September 8 and 10, 2026, suspected Chinese-linked attackers used Cairn to discover a previously unknown zero-day vulnerability in a medical imaging application. The attackers, targeting mainly East Asian educational institutions outside China, simply instructed the AI agents to “obtain shell access,” leaving the autonomous system to independently discover and exploit the flaw. The campaign was exposed when the attackers left their Cairn instance publicly accessible without authentication, inadvertently revealing their operations.
Cairn is a legitimate open-source penetration-testing framework designed to coordinate multiple AI agents for research and defense strengthening. However, its adoption by malicious actors highlights a significant shift in the cyber threat landscape. Unlike the targeted, human-guided exploitation seen in the Armory Castle campaign, the use of Cairn represents an attempt to automate vulnerability discovery and exploitation. Yet, the efficacy of these autonomous agents is not absolute. Amazon researchers noted that the AI agents were not always successful; for instance, a 16-hour, 148-action effort to take control of a Redis server failed. This distinction is critical: while AI tools can accelerate certain stages of an attack, they are not infallible and can still encounter technical barriers that hinder complete system takeover.

The convergence of these two narratives—one involving a state-sponsored group exploiting a specific, high-value zero-day with custom tooling, and the other involving the misuse of autonomous AI for broader reconnaissance—underscores the dual challenges facing defenders. The VMware incident demonstrates the persistent threat of targeted espionage campaigns that require rapid patch management and network segmentation, particularly for critical infrastructure like virtualization servers. The Cairo incident, meanwhile, signals the emerging risk of AI-driven attacks that may bypass traditional signature-based defenses by adapting to new environments in real-time, albeit with limitations.
For organizations, the implications are clear. The exploitation of CVE-2026-123890 emphasizes the need to monitor for unauthorized SSH access and anomalous API calls to vCenter management interfaces. Additionally, the use of tools like BONESAW and BARECLAW indicates that attackers are deploying multi-stage payloads to maintain persistence and evade detection. As AI tools like Cairn become more prevalent, defenders must consider the potential for automated, low-fidelity attacks that may serve as precursors to more sophisticated intrusions. The unresolved technical issue remains how to effectively detect and mitigate the hybrid threat of human-directed espionage augmented by autonomous AI capabilities, a challenge that will likely define the next phase of cybersecurity strategy.



