Global – October 31, 2026 -- CrowdStrike's Falcon Guardian now traces supported AI agent sessions from prompt to impact on the endpoint, scoring every agent from 0 to 100 and blocking dangerous actions at runtime.
The product targets a blind spot in network-based monitoring. A gateway sees the conversation, and a web gateway sees an agent connect to a trusted site. Neither sees what happens inside the machine, where a hidden instruction can become a shell command or a downloaded file can run with the user's access.
Agent Graph links the AI layer to the execution layer
Falcon Guardian runs on the existing CrowdStrike Falcon sensor, which already sees processes, files and network connections. It now also sees prompts, models, tools and skills. CrowdStrike Agent Graph connects each agent to the users, hosts, models, prompts, tools, skills, MCP servers, processes, files and detections it touches, and keeps those links current as agents act.
Three Falcon Guardian views read from the same graph
Agent inventory: The view lists sanctioned and shadow agents across the fleet, grouped by type. Coding agents, desktop assistants and command-line agents from Anthropic, OpenAI, Cursor and others appear side by side. In CrowdStrike's example, a single Claude Code agent has one user, three installs, eight tools and 12 sessions.
Session graph: The view orders the causal chain from the user and host to the prompt, each tool call, and the resulting processes, files and network connections. In CrowdStrike's example, a developer pastes a setup note from an untrusted source into Claude Code. The prompt looks harmless from the network. The graph shows three tool calls: the agent contacts an outside host, installs a package and runs an installer from Downloads. The detection attaches to that final tool call. Falcon Guardian blocked the process and quarantined the file. Cloud keys and personal data pasted later in the session were flagged separately on the same graph.
Agent threat score: Each agent carries a 0-100 score built from detections tied to its sessions. Agents are colored by score, so teams can triage the highest-risk agents first instead of working the oldest alert. One click on a critical agent opens the sessions and detections behind its score.
Because the three views share one graph, analysts can move from fleet to agent to session to the exact triggering step without switching tools or rebuilding a timeline.
Agent hooks add prompt text and pre-execution enforcement
For supported agents, Falcon Guardian agent hooks plug into an agent's native hooks system. They capture activity when a prompt is submitted, before a tool call runs and after it completes. This adds exact prompt text, tool inputs and tool outputs to the graph alongside the sensor's view of what executed. The sensor can also see the request and the LLM response through endpoint network inspection.
Hooks also serve as an enforcement point. Falcon Guardian can check a prompt or tool call against policy and stop a dangerous action before it executes. Agent hooks are available today for Claude Code, OpenAI Codex, GitHub Copilot, OpenCode and Cursor.
CrowdStrike says it is the only vendor combining native endpoint visibility, endpoint network inspection and agent-level controls to stop malicious agent behavior at runtime.