Approval Fatigue and Human-in-the-Loop Security

The MCP Security Risks You Can't Afford to Ignore - Ariel Shiftan

Topic 10 of 12

Published Mar 2, 2026 · Clip 59:36–1:05:26 (5:50)

Ariel discusses the tension between allowing agent access and requiring runtime user approval, drawing parallels to Android permission prompts that users blindly approve and to his work at Magic Leap. He advocates designing platforms with graded access levels so developers naturally choose the safe option and users only approve genuinely sensitive actions, suggesting coding agents should sandbox and classify command safety so humans only weigh in on real risks.