OpenAI dot AI agent consumes high token volumes autonomously

Testing by AI security and governance firm Geordie has revealed that OpenAI’s “always-on” personal AI agent, dot, can consume significant amounts of tokens without direct user prompting. In one test, a single dot agent generated 155 separate activity sequences and consumed over 300,000 tokens in a single day without any human interaction.

Geordie Chief AI Officer Hanah-Marie Darley noted that dot behaves differently from other personal agents, such as Meta Muse or xAI Grok Bot, due to its high level of autonomous activity. Even with only read-only permissions and no connection to personal communication tools like email or Slack, the agent entered a research mode to survey systems and collect information, creating daily journal-like logs.

The agent operates by spinning up multiple sub-agents; for example, one might check code repositories while another inventories collected data. This autonomous behavior poses potential cost risks for enterprises, as Darley estimated that running the agent on the OpenAI Astra model could cost between $5 and $10 per day in token usage even when the user is not actively interacting with it.