OpenAI's dot agent consumed over 300,000 tokens autonomously in one day

Security firm Geordie tested OpenAI's always-on assistant dot with read-only permissions and no connections to personal apps. The agent generated 155 activity traces and used more than 300,000 tokens in a day without further human prompts. Geordie's chief AI officer said dot was unusually active, spinning up subagents on its own and surveying available systems.

Geordie began evaluating OpenAI’s continuously available dot after its September launch, placing it alongside Meta Muse and Grok Bot. In tests led by chief AI officer Hanah-Marie Darley, dot received only read-only access and was kept away from personal apps, email, and services such as Slack. It generated 155 activity sequences and exceeded 300,000 tokens in a single day without additional prompts.

Darley described dot as more proactive than comparable personal agents. It surveyed available systems, gathered context, maintained something like a daily journal, and checked assigned tasks. Notably, it launched subagents on its own rather than responding to scheduled or user-driven requests. Dot runs at least two agents: one for user interaction and another continuously in the background.