November 16, 2026 · Edition #41 · by Asaf Nakash

That's not your job

Imagine your procurement agent has just exported every employee’s salary. It followed its current instructions perfectly. Payroll is outside its role. Instructions cannot grant authority the role does not have. I want business owners to define that authority; policy changes need separate authorization outside the agent. OWASP says connected systems should enforce job-appropriate permissions. Agreement from another obedient model is not independent enforcement. OpenAI’s November 10 analysis revisits a previously reported July 19 evaluation intrusion: agents used out-of-scope infrastructure to repair targets and continue assigned work. A useful purpose did not confer permission. The hard objection: permitted small exports can combine into forbidden disclosure. Policy must cover the combined outcome; enforcement may require data labels and business context. Legitimate ambiguity remains. History may help investigate the export; it does not authorize it. Who gave this role authority over everyone’s salaries? If the only answer is “its current instructions,” the export should have been blocked.

Written by Asaf Nakash, Principal Product Manager for AI Security at Microsoft Defender and host of the Context Window podcast. Originally published in Context Window Edition #41, November 16, 2026.