DTEX Finds Claude Cowork Could Enable Rapid Data Exfiltration Through Business Apps

· Original article ↗

Summary

DTEX researchers found that Claude Cowork could access and transfer sensitive data from business apps and endpoint files using simple prompts. They warn that weak logging, monitoring and access controls can make AI-driven insider threats harder to detect.

Key points

  • DTEX tested Claude Cowork workflows that summarized Salesforce information into an Outlook draft and transferred selected files through the Cowork app.
  • Researchers said the tests took 10–30 minutes of preparation and used simple, single-turn prompts.
  • The agent could access SharePoint, OneDrive, Outlook, Salesforce data and files on the user’s endpoint, according to DTEX’s tests.
  • Dispatch relays commands from a user’s phone to a desktop Claude agent; the article also describes plugins for business applications.
  • DTEX framed the risk as an IT governance and visibility problem, not exploitation of a software bug or configuration vulnerability.
  • The researchers warned that inadequate prompt logging, access controls and monitoring can hinder detection and investigation of AI-related data leakage.

Article Details

Event Type
Research demonstration of insider-threat risks from AI agents integrated into corporate workflows
Impact
DTEX researchers found that Claude Cowork could access sensitive corporate data across connected systems and, in tested scenarios, prepare data for exfiltration in 10–30 minutes using simple prompts. The risk arises from broad access and inadequate logging, monitoring, and access controls, not a software bug or configuration vulnerability.

People

Vendors

Products

Industries

Related Articles