OpenAI’s dot assistant generated 155 traces without prompts in a day
Geordie found that an OpenAI dot assistant generated 155 activity traces and used more than 300,000 tokens in a day without further human prompts. The finding highlights how personal AI assistants can work proactively in the background, raising questions about how much activity and access users expect from them.
In testing, Geordie’s chief AI officer gave dot read-only permissions and did not connect personal apps, email or messaging services. At least 50 traces occurred within two minutes, and most happened over a few hours; dot also launched short-lived sub-agents to inspect available information, including code repositories. The article describes dot as always running at least two agents, one interacting with the user and another gathering data in the background.
- Dot used over 300,000 tokens in one day without prompts.
- Geordie recorded 155 activity traces during its test.
- The assistant launched sub-agents while given read-only access.
New here? Start with this
Personal AI assistants are designed to help with tasks by using information a person makes available to them. Some can also run in the background, checking information or carrying out steps without waiting for a new instruction each time.
OpenAI’s dot is an assistant that can use separate AI agents for different tasks, including one that interacts with the user and another that gathers information. In this test, it had read-only access and was not connected to personal apps, email or messaging services; it inspected information that was available, including code repositories.
The issue matters because background activity can consume computing resources and may do more than users expect, even when an assistant cannot change or send information. Understanding what an assistant can access, how often it acts and what it does in the background helps people judge whether its behaviour matches their expectations.
Both sides, in good faith
The strongest fair case each way — we don't pick a winner.
The case for
Proactive background work could make personal assistants more useful by gathering relevant context before a user asks, reducing repetitive prompts and helping with complex tasks. In this test, dot had read-only access and no personal apps connected, so the activity may illustrate a design aimed at preparation rather than unauthorised changes.
The case against
A user may reasonably expect an assistant to act only when prompted, even if its access is read-only: frequent traces and substantial token use can consume resources and make it harder to understand what the system is doing. Running agents in the background also raises questions about privacy, oversight and whether users have clear controls over the scope and pace of activity.
Read the full article at the source →
Originally published by The Register as “OpenAI’s dot assistant burned 300K+ tokens in a day without a single prompt”.