Your AI agent could become your biggest insider threat 

DTEX researchers demonstrated that Anthropic's Claude Cowork agent tools (including Dispatch and Salesforce plugins) can be instructed with simple prompts to access and exfiltrate corporate data—SharePoint, OneDrive, Outlook, Salesforce and endpoint files—within 10–30 minutes; the issue is framed as an IT governance/visibility and insider-threat problem rather than a software vulnerability, highlighting the need for access controls, prompt auditing, and monitoring.