Risk-contingent autonomy boosts trust in LLM agents

tianshi_li · x · 2026-08-18

A paper accepted to COLM explores personalization and privacy risks in LLM agents. It proposes "risk-contingent autonomy," where agents delegate control back to users upon detecting potential privacy leaks. Experiments show this design enhances perceived control, mitigating privacy concerns and trust loss from personalization, similar to Claude Code's auto mode.

Original post →

More from Safety

Safety channel →