Testing Two New AI Agents by Handing Over Passwords for Seven Days
A journalist spent a week testing Meta's Muse and OpenAI's Dots, two new AI agents, by granting them access to passwords and personal tasks. The agents can handle purchases, emails, forms, and proactive research, and are designed to feel conversational. Computer scientists helped evaluate how much trust users should place in them.
Meta launched Muse in September, an agent with its own computing space that can buy items, email and complete forms when granted access. OpenAI followed weeks later with dots at DevDay 2026, aimed at top-tier subscribers and capable of ongoing background research. Both mimic friendly texting, with names and avatars.
During a weeklong trial tied to an apartment move, the writer used Muse to search Facebook Marketplace and message sellers, successfully obtaining a dining table. Giving dots the same Facebook login triggered Meta's bot detection, temporarily locking Instagram, Facebook and WhatsApp. Dartmouth's Nikhil Singh notes the underlying approach has existed for years; adoption scale may matter most.
If such agents become mainstream, people may delegate shopping, scheduling, correspondence, and research, potentially saving time but also exposing sensitive accounts and spending. Platforms could struggle to distinguish helpful automation from abuse, as the Facebook lockout suggests. Less technical users may be especially affected, since these products are designed to feel conversational and require little training. Researchers' caution suggests users may need to weigh convenience against access.