Bi.Zone: Up to 72% of organizations show signs of AI agent usage in their infrastructure

Bi.Zone reported that up to 72% of organizations show signs of using AI agents in their infrastructure, based on SOC telemetry. The most prevalent agents installed are Cursor (24%), Codex (24%), Copilot (17%), and Claude Code (8%), but presence does not mean active use. When measuring actual activity, the top five agents are Codex (54%), Claude Code (13%), Cursor (12%), Zed (9%), and OpenCode (9%). This discrepancy indicates that tool prevalence does not always reflect real demand or risk. AI agents can autonomously perform actions, interact with external services, and access corporate resources such as source code repositories, cloud platforms, internal APIs, and documents. If compromised, attackers could exploit these capabilities. In one investigation, Bi.Zone SOC found evidence of a workstation compromise where a user used an AI agent to launch tasks on other computers; the agent installed offensive tools and executed commands on behalf of the user. Such scenarios complicate security work because analysts must determine which actions were user-initiated versus agent-autonomous. A separate risk is shadow AI—employees using AI tools, plugins, or agent components without IT and security approval. Unlike regular software, these solutions can make decisions and access corporate data, so they require special oversight. Teymur Kheirkhabarov, product director at Bi.Zone, emphasized that AI agents form a new class of corporate tools needing systematic control due to access to internal data, repositories, and APIs; misconfigurations or overprivileges could lead to serious incidents. He noted the importance of analyzing technical indicators to understand actual usage, which surveys or software inventories cannot provide. Comprehensive telemetry analysis enables inventorying AI agents and their components, but the choice of management approach—restriction or controlled adoption—remains with security directors.