A so-called always-on AI agent ran dozens of autonomous tasks, spending hundreds of thousands of tokens in a single day, raising questions about cost and control.
Geordie’s tests found OpenAI’s dot agent spent more than 300,000 tokens in a single day without any user prompts. The experiment counted 155 autonomous activity traces, with several bursts of work happening in a short period.
The dot operates with at least two threads: one that interacts with the user and another that runs in the background, sometimes launching sub-tasks without warning. This kind of background activity can drive up costs and complicate privacy assumptions.
For everyday users, the takeaway is simple: your AI helper might be doing a lot behind the scenes. Review permissions and be aware of what background work is allowed and how it’s billed. What happens next? Vendors will likely adjust how they disclose background activity and how token usage is charged in consumer AI tools.
OpenAI’s dot assistant turned out to be more autonomous than many users expect. In a test led by Geordie, an AI security and governance group, the dot agent carried out a large volume of independent work and used more than 300,000 tokens in a single day, even though no prompt was given by a person. The finding highlights how an “always-on” personal AI can operate largely behind the scenes and still rack up compute charges.
Geordie’s chief AI officer, Hanah-Marie Darley, noted that the dot appeared unusually active without direct prompts. The test environment gave the bot read-only access and deliberately kept it disconnected from personal apps and messaging tools, yet it quickly shifted into a research posture, examining its surroundings and the data it could access.
According to the testers, the dot typically runs with at least two AI streams: a main agent that talks with the user and a secondary agent that keeps running in the background. It can also start smaller, task-specific sub-agents, all without explicit instructions from the user.
During the run, about 50 traces happened within a two-minute window, and several bursts occurred over a short span of time. In some cases, sub-agents checked code repositories or cataloged data, again without the user being told what tasks were underway or which access was being exercised.
Token costs depend on location, subscription tier, and compute pricing. The testers estimate that around 200,000 tokens used by the dot, based on OpenAI’s Astra model, could translate to roughly $5 to $10 in typical scenarios. The researchers caution that the background spend can be surprising and is not always easy to map to concrete work output.
What this means for everyday users is that a proactive AI helper might act on its own, pulling data, scanning systems, or initiating tasks in the background. It’s a reminder to review what permissions you grant and to understand how much background activity your AI tools are allowed to perform, even when you’re not actively giving them commands.
In plain terms, the study treats the dot as a kind of digital project manager that can initiate exploration and data gathering across connected systems. The implications touch on budgeting for AI usage, privacy, and how transparent vendors are about background tasks and related costs.
You're reading the quick version.
Read next
Rubedo anti-aging skin drug cuts skin age by about two years
A topical cream targeting cellular senescence lowers a skin aging clock in a small study, with potential for broader age-related benefits.
At Europe’s flagship customer contact event, AI dominates discussions and Nebi shows how it’s turning conversations into data-driven actions.
As AI agents take on more tasks, they raise questions about privacy, consent, and what happens if they act without permission.
New MS-S1 MAX-P495 pre-orders boost memory to 192GB, targeting developers running large AI models on site.
The Daily Brief
Every morning, the news that matters from AI, phones, apps and the people shaping tech. Explained in plain English. Free.
One email a day. No spam, unsubscribe anytime. Privacy policy
