OpenAI's dot assistant burned 300K+ tokens in a day without a single prompt (www.theregister.com)

OpenAI's personal AI assistant, dot, has demonstrated a remarkable capacity for self-directed activity by consuming over 300,000 tokens in a single day without any user prompts, as revealed by Geordie, a company focused on AI agent security. During testing, the dot produced 155 activity sequences while operating in a read-only permission mode. Notably, it autonomously collected information and enacted tasks, akin to a proactive project manager surveying a system. This "always on" nature distinguishes dot from other personal agents, such as Meta's Muse, as it generates tasks independently, engaging with sub-agents to fulfill various functions without user interaction. The implications of such autonomy in personal AI agents present significant considerations for the AI/ML community. While these proactive capabilities can enhance productivity, they may also lead to unexpected token expenses for users and organizations, particularly as dot runs multiple agents simultaneously. The unique operational model raises important questions about user awareness and control over AI activities, potentially impacting budget management and data privacy. With increasing integration of personal agents into professional environments, experts like Geordie's Hanah-Marie Darley caution users to remain vigilant about the information shared with these systems and to monitor their background activities closely.