Anthropic Adds EDR Visibility Into AI Agent Activity Inside Claude Cowork VMs
Written by AI from the linked story. Not reviewed by an editor. How this works
The launch of EDR sensor support in Claude Cowork highlights a previously overlooked security blind spot: AI agents executing shell commands and code inside local virtual machines were entirely invisible to host-level security tooling. Without this visibility, security teams had no way to detect malicious, anomalous, or policy-violating behavior initiated by AI agents on behalf of users. This matters because AI agents are increasingly granted broad system-level permissions, making them a high-value target for misuse or compromise. The absence of monitoring inside sandboxed AI execution environments represents a systemic gap as agentic AI use expands across enterprises.
Tactical Insight
Immediate actions
- Enable EDR sensor integration within Claude Cowork on all Team and Enterprise plan deployments for macOS and Windows immediately.
- Audit current AI agent deployments to identify any execution environments that lack endpoint-level monitoring coverage.
Long-term improvements
- Establish a formal policy requiring that all AI agent runtime environments support EDR or equivalent telemetry before being approved for enterprise use.
- Maintain a living inventory of all AI tools and the system permissions they are granted, reviewing it on a quarterly basis.
- Work with AI vendors to ensure security visibility is a contractual and technical requirement in procurement processes.
Detection measures
- Configure EDR alerts specifically for unusual shell commands or code execution patterns originating from AI agent processes.
- Route AI agent execution telemetry into your SIEM for correlation with broader threat detection rules and anomaly baselines.