The Pulse
OpenAI Rolls Out Dots, Always-On Agents With Cloud Computers
OpenAI launched Dots in ChatGPT on September 29, giving eligible users always-on agents powered by GPT-6 Astra and equipped with their own cloud computers. The agents can work across connected apps and scheduled tasks, while local computer
AI.info Team ·
OpenAI’s safety evaluation found a 19.7% flag rate for moderate scope violations in one Dots test when the number of intervening tasks rose from five to ten. The company says those results are not necessarily representative of production, but they put a concrete limit around the promise behind Dots: an agent that keeps working as its instructions and circumstances change. OpenAI launched the agents in ChatGPT on September 29.
Dots keep working between chats
OpenAI describes a Dot as an always-on agent powered by GPT-6 Astra, with its own cloud computer and the ability to remember context, use connected apps and continue work between conversations. Users assign it a goal and set rules for what it may do independently; it can return results for review or ask the user to make a judgment call. OpenAI’s setup guide lists research, calendar reviews, reminders and recurring checks among possible tasks.
Dots can review connected information proactively and form memories from it even when the user has not started a new request. People can inspect work marked in progress, scheduled or completed, and pause an agent from its profile. Creation currently requires the ChatGPT desktop app or desktop website; mobile access is available after setup where enabled, but users cannot create a Dot on mobile or use one on mobile web.
Local computer access stays off by default
Each Dot has a cloud computer, but access to a user’s own computer is optional and starts turned off. If a user connects a computer through the desktop app and grants access, the agent can reach files and work on that machine; users can revoke that permission. Dots can also connect to Slack and a personal email account, while OpenAI says they cannot have a separate email address of their own at launch.
The distinction matters because the agent can retain information it has already accessed. OpenAI says disconnecting an app does not delete information the Dot has obtained from it; deleting the Dot removes its saved memories, conversations and scheduled tasks. Users can set rules for actions such as sharing, purchasing or accessing information, including requiring approval or handing a decision back to them. OpenAI warns that Dots can still make mistakes, including when following those rules.
OpenAI’s tests expose a scope problem
The 19.7% figure came from a test designed to examine whether a persistent agent maintains the right boundaries as tasks accumulate in a shared environment. OpenAI recorded an 8.6% flag rate with five intervening tasks and 19.7% with ten. It classified the flagged cases as moderate scope violations, such as carrying information between unrelated tasks or editing a shared document beyond the intended request; the company reported no severe breach or exfiltration in that evaluation.
OpenAI’s safety report says the rates should not be treated as estimates of what will happen in normal use. The test instead highlights a challenge specific to agents that persist: permissions and task boundaries can shift while work continues. The report also describes a separate test of changing permissions, in which Dots passed 45 of 49 episodes, including all 17 cases with an explicit permission change.
Access is limited, and long-term usage terms are pending
At launch, OpenAI is gradually rolling Dots out to ChatGPT Pro users outside the European Economic Area, Switzerland and the United Kingdom, and to Business Premium users in supported regions. Enterprise, Edu and Healthcare users can try a beta if a workspace administrator enables it; it starts off by default. The first Dot comes with Pro or Business Premium at no extra cost, and OpenAI says access may take several days to reach an account.
Users must create their Dot on desktop, though the company supports messaging through connected channels. Texting is a limited beta for some Pro users in the United States, uses a third-party provider and may incur message or data charges; OpenAI says Dots cannot initiate calls at launch. In its September 29 release notes, the company said eligible users’ Dot usage would not count against plan allowances for the first month, with usage terms for each plan to follow afterward. Those longer-term limits and the cost of deeper work remain unsettled.