October 22, 2026
Falcon Guardian Overview
See how CrowdStrike Falcon Guardian finds, controls and protects AI agents on the endpoint. It uses the Falcon sensor you already have, with no hooks and no SDK.
This video was created and contributed by, CrowdStrike.
Overview
CrowdStrike Falcon Guardian secures AI agents where they actually run: on the endpoint, at runtime. It uses the existing Falcon sensor, so there are no hooks or SDK to deploy. Guardian combines AI telemetry and operating system telemetry, from the prompt through to the process, across your entire AI estate. One Fortune 500 company found 18,000 AI agents on its endpoints and had approved only 300. Guardian discovers every agent on day one, including shadow AI that is dormant or hiding. It then blocks unapproved apps, masks sensitive data such as credit card numbers before it reaches the model, and stops attacks like indirect prompt injection and malicious MCP servers before credentials leave the machine. A single causal timeline and agent graph tie prompts, skills and tool calls to process spawns, file reads and network egress. Security teams can investigate with the Falcon MCP server, and real-time token usage by user, agent and model gives finance the same source of truth as security.
Key points
- Full AI visibility: Discovers every agent on day one, including shadow AI, through sensor scans and DNS queries. The demo's Fortune 500 customer had 18,000 agents running and 300 approved.
- Policy and data protection: Blocks unapproved tools, steers users to the approved app, and masks sensitive data like credit card numbers at the prompt.
- Layered attack defense: Stops prompt-layer attacks such as hidden indirect prompt injection. Catches process-layer attacks such as a malicious MCP server that looks clean to the prompt layer.
- One causal timeline: The agent graph combines prompts, skills and tool calls with process, file and network activity, and scores risk per agent.
- Faster investigation: Analysts can query the Falcon platform with the Falcon MCP server to find where else a malicious skill ran and quarantine the affected hosts.
- Cost and usage tracking: Real-time token usage by user, agent and model gives finance spend visibility and security anomaly detection.