1-minute read of a 4 min report

Sophos Urges Rules at Sensitive Uploads Before AI Judges Intent

By

1-minute read

What a defender needs to know

A model watching an employee copy a customer list may describe the transfer accurately and still misread the purpose. Sophos researcher Nash Borges argues that endpoint intent scoring works best alongside explicit policy at sensitive actions, such as an upload, share or payment. A person using a job-relevant file may look normal even when the transfer is harmful.

For a customer-data upload to an external AI service, map the destination, resource and policy decision first. Run job-fit scoring in observation mode. Preserve the action, resource, destination, policy result, score and session context for reviewers. Set a daily review budget and measure both suspicious actions surfaced within it and legitimate work that a proposed prompt or hold would interrupt. Test a job-relevant harmful access mixed with ordinary activity; a low anomaly score must not override a rule against sending customer data to a prohibited destination.

Promote the model-based intervention when a prohibited upload hits explicit policy and an uncertain permitted transfer leaves a reviewable record within the team’s preset review and interruption budget. Borges evaluates this control design for endpoint teams.

Primary sources