CrowdStrike Adds Endpoint Visibility for AI Agents That Open Files, Run Code and Use Credentials

AI agents now do far more than answer questions. According to CrowdStrike, they open files, use credentials, run code, install software and run models locally, all at machine speed. Every one of those actions happens on the operating system, which is where attacks land. CrowdStrike argues that monitoring only from the network leaves a gap: a gateway sees the conversation, and a web gateway sees a connection to a trusted site, but neither sees what happens inside the machine. In that gap, a hidden instruction could become a shell command, or a downloaded file could run with the user's access.

CrowdStrike's answer is Agent Graph, which sits underneath its Falcon Guardian product and runs on the Falcon sensor that customers already have. The company says it connects the AI layer (prompts, models, tools and skills) to the execution layer (processes, files and network connections) in one chain, tracing supported agent sessions from prompt to impact. It links each agent to the users, hosts, models, tools, MCP servers and detections it touches.

Three Falcon Guardian views are built on it: agent inventory, session graph and agent threat score. The inventory is designed to show AI agents across a fleet, whether sanctioned or shadow, including coding agents, desktop assistants and command-line agents from vendors such as Anthropic, OpenAI and Cursor. For supported agents, it shows how far each one reaches. This summary covers only the opening of the article.