I Think I Found an AI Agent Worth the Risk

A new class of AI agents that manage your digital life is emerging, and they are starting to get things done despite significant privacy risks.

The AI agent Instinct, which operates via iMessage and WhatsApp, has gained significant attention in the Bay Area. The company behind Instinct, also named Instinct, is reportedly seeking to raise $1 billion at a $10 billion valuation, adding to its existing $350 million in funding. This surge in interest coincides with Meta's AI assistant, Muse, becoming the top free app in the Apple App Store with over 900,000 downloads, despite a critical security flaw that could grant attackers full control over a user's Mac. Instinct's appeal lies in its user-friendly interface, integrating with email, calendar, and messaging apps to perform tasks like booking reservations and managing travel logistics. However, both Instinct and Muse present significant security and privacy concerns, with reports of data retention, excessive API usage, and potential phishing vulnerabilities.

The emergence of AI agents like Instinct and Meta's Muse highlights a growing divide in user adoption of AI tools, driven by the perceived value proposition of task automation. For those who see agents as productivity boosters, the infrastructure costs of AI are acceptable; for others, who view them as enhanced search tools, the trade-off is less compelling. This polarization is further exacerbated by the current state of AI agents, which often require significant user guidance to be effective, potentially leaving less proactive users behind.

Instinct's success, particularly its integration through familiar messaging platforms and its ability to handle complex tasks like travel rebookings, suggests a potential 'form factor' breakthrough for AI agents. This user-centric approach, avoiding complex interfaces for simple text interactions and proactive suggestions, differentiates it from less intuitive competitors. The reported security concerns, however, remain a significant hurdle, with user data potentially being retained and models trained on conversations, presenting a clear risk-reward calculation for users.