OpenAI's dot assistant burned 300K+ tokens in a day without a single prompt

OpenAI's always-on agentic assistant, dot, consumed over 300,000 tokens in a single day without any human interaction. This was discovered during testing by AI security and governance group Geordie, which found that dot was actively researching systems, gathering data, and even spinning up sub-agents without prompting. Compared to other personal agents, dot stands out for its proactive behavior, constantly checking on its own tasks and collecting information about the user's environment.

The testing revealed that dot operated with two agents at all times - the primary agent interacting with the user and the secondary agent continuously running in the background. This behavior led to a high number of agent traces, with at least 50 happening within a two-minute period and many occurring within a few hours. While the token usage can vary based on factors like user location and subscription tier, the cost of 307,000 tokens for testing was equivalent to around $5-$10.

This constant, background operation of dot raises concerns about unexpected AI spend, especially for users who may not realize how much their agents are doing without their knowledge.