Not only are AI agents all the rage in the tech world, but they’re becoming more mainstream by the week. Apple recently upgraded iOS and macOS with Siri AI, while earlier this month, Meta also launched “Muse,” an agent that can perform tasks on your behalf—and potentially expose your private information. Now, it’s OpenAI’s turn to launch an AI agent, but its take on the trend is a bit different, and less accessible, at least at launch. The company is calling it “Dots.”
What are OpenAI “dots?”
“Dots” isn’t a single AI agent platform. Rather than having all OpenAI users tapping into the same “Dots” system, users create their own individual “dot” for a specific purpose. OpenAI says dots are always-on agents powered by the company’s GPT-6 Astra model that learn over time based on your responses. They can connect to over 4,000 apps at launch, including ChatGPT (of course), Slack, and Teams.
According to OpenAI’s pitch, you won’t always have to instruct your dot to do things for you; instead, the agent will take anticipatory action, guessing what you need done based on previous experience. That might mean identifying a bug in the app you’re developing and deploying a fix without being asked. One early dots tester reportedly forgot to file an invoice for work, and their dot prepared one on his behalf, only sending it after gaining his go-ahead.
Dots run on their own cloud computer. OpenAI says you can view this computer at any time to check what it’s up to and what it has done. Dots can connect to other devices, including your laptop, if you grant it permission first. OpenAI says you can continue adding to its workload without worrying about slowing your bot down: It can supposedly handle many concurrent tasks. Maybe you’re working on an app, and you task the dot with monitoring customer feedback and fixing small issues with code. Perhaps you’re a content creator, and your dot analyzes interview transcripts for clippable segments, while drafting up social media posts.
Like other AI agents, dots are chat-based. The idea is to simulate the experience of texting a friend or colleague. You text them in ChatGPT, Slack, or Teams, and they text you back with updates and questions. They can even text you first when there’s been a development. At launch, OpenAI says you can create a “primary” dot. You can give it a name and “make it your own,” though it’s not clear exactly what that means.
Are OpenAI’s dots safe to use?
OpenAI says that dots build on ChatGPT’s security systems “with additional safety and privacy safeguards.” The company says all dots work on their own cloud computer, separate from your own machine—unless you choose to connect the two. The company claims that dots can sign you into accounts using saved passwords without actually exposing those passwords to the AI model. In fact, when you sign into an account, the AI model pauses working until sign-in is complete. Your dot will then resume its activity.
Dots will work in the background as part of its “proactive research” feature, but it will only connect to apps you’ve already permitted it to, using tools that cannot make changes. While prompt injection is always a risk with agentic AI, OpenAI says that dots have built-in safeguards to protect against malicious instructions. If the AI detects something suspicious, it will stop the dot from proceeding. In addition, you can set up custom rules to ensure the bot will never perform a task you don’t want it to, such as sending emails.
Dots also operate on a system called “Auto-review,” which compares upcoming steps against your instructions, rules, and safety requirements. If it detects an issue, it stops itself. The dot can then investigate further on its own or come to you with the problem. However, OpenAI acknowledges that dots can still make mistakes.
Who can use dots?
Unlike Meta, which is actively pushing Muse on all users, OpenAI’s dots are not as widely accessible as of this writing. At launch, dots are only available to paying users, including ChatGPT Pro, Business Premium, and Enterprise users. What’s more, you can only set up one dot right now, though OpenAI says you’ll be able to set up multiple dots in the future.