OpenAI Dots and Meta Muse offer privacy toggles, but gaps persist
Meta's Muse trains on your data by default and Meta can still see inside it. OpenAI's Dots promises tighter rules. Here's how the two AI agents really compare on privacy.
OpenAI unveiled Dots, its always-on AI agents, at DevDay in late September and pitched them as the privacy-safe rival to Meta's Muse.
Muse lets Meta train models on your inputs unless you opt out, and Meta can still access the data in each user's private virtual computer.
Muse has already shared one user's home address with a Marketplace stranger, and a zero-day flaw that could hijack it was found and patched.
Dots ask for approval before purchases and other sensitive steps, but they inherit your ChatGPT training setting and need the $100-a-month Pro plan.
Quick read · 2 min
OpenAI and Meta are fighting over who you can trust with an AI agent. OpenAI's Dots, unveiled at DevDay in late September, are always-on helpers that run on GPT-6 Astra, connect to thousands of apps and ask for approval before purchases. OpenAI pitched them as safer than Meta's Muse.
Muse has had a rough start on privacy. It trains on your inputs unless you opt out, Meta can still access the data in each user's private cloud computer, and it once gave a user's home address to a stranger on Marketplace. A zero-day flaw that could hijack it has been patched.
Muse: turn off model training in its data controls.
Dots: your ChatGPT training setting carries over.
Both: skip extra permissions and don't store passwords.
Dots costs $100 a month on ChatGPT Pro. Meta says a cryptographic lock on Muse data is coming later this year.
One AI agent handed a user’s home address to a stranger on Facebook Marketplace and invited them over. Another, its maker swears, would never be so careless. That is the pitch war between Meta’s Muse and OpenAI’s new Dots, and if you’re thinking of letting either one run errands for you, AI agent privacy is the question that matters most.
These aren’t chatbots you ask one question. They’re agents: software that stays switched on, reads your email, calendar and messages, browses the web and acts for you. To be useful they need far more of your personal data than any app before them. Both companies say they’ve built in protections. The record so far shows how much depends on settings most people never open.
01
What OpenAI promised with Dots
OpenAI introduced Dots at its DevDay event in San Francisco in late September, with CEO Sam Altman promising that Dots would raise the bar for privacy among top AI labs. Dots run on the company’s GPT-6 Astra model, show up as cute, customizable blobs and work around the clock on tasks you hand them. Each Dot gets its own cloud computer and browser, and OpenAI says its plugins connect to more than 4,000 apps. You can talk to one through ChatGPT, Slack or Microsoft Teams.
The safety design rests on a few rules. OpenAI says the background research a Dot does on its own uses read-only tools, so it can’t send messages, change your connected apps or drive a browser while it’s poking around. Purchases need your approval, and changing passwords stays your job. A Custom Rules tool lets you draw lines, for example telling your Dot never to spend over a set dollar amount, which Altman showed on stage.
Businesses get more. OpenAI pitched enterprise customers stronger data controls and the option of zero data retention, meaning nothing is kept on OpenAI’s servers. One OpenAI executive also took a swipe at Meta, pointing to the 1.2 billion users of OpenAI’s own AI product and saying the company would take care not to ship something that makes Muse-style mistakes.
02
The Meta Muse privacy problems so far
Meta’s Muse arrived a couple of months before Dots, marketed by Mark Zuckerberg as built from the ground up for privacy and security. Each user’s data sits on what Meta calls a secure virtual machine, essentially an isolated Linux computer in the cloud with its own browser and storage. Meta said most of its work went into careful design so the agent operates more safely, while admitting Muse will still make mistakes.
The promise and the practice haven’t matched. Your data is walled off from other users, but not from Meta itself. The company says it plans to add a way to cryptographically and verifiably stop itself from reading your virtual machine later this year, which means it can read it today. And Muse lets Meta train its models on what you put into it by default.
The incidents stack up:
A security researcher found a zero-day flaw, a bug unknown to the maker, that could let someone take control of Muse. Meta has since patched it.
A report says several serious security problems surfaced right before launch, including one that could have let users reach Meta’s own internal databases.
A journalist said Muse uploaded and read his private messages without being asked, and a YouTuber said it offered his address to a Marketplace buyer. In both cases the agent seemed to be working as designed, but went further than the user expected.
Meta rolled back an internal test that used human contractors to handle calls after its own employees raised privacy worries.
None of this stopped Muse from topping the App Store and reaching 600,000 daily US users within weeks, according to app-tracking firm Apptopia.
Muse is built from the ground up for privacy and security. Your data and credentials live on the Muse Secure VM — an isolated linux computer with a browser, CPU, memory, and storage.
Dots or Muse: which AI agent is better for privacy?
On paper, Dots comes out ahead, mostly because it has fewer scandals and clearer approval steps. But there’s a catch: Dots launched only for the ChatGPT Pro tier at $100 a month, plus some business plans, so far fewer people are using it and fewer problems have had a chance to surface. A smaller, paying audience is easier to protect than a free app aimed at billions.
There’s also a shared weakness. Any agent that reads your inbox can be tricked by a booby-trapped web page or email into doing something you didn’t want, and both companies say only that their safeguards reduce the risk. And neither product changes the basic trade: the agent is only helpful because it knows a lot about you. Some testers have balked at simply typing bank details into a chat when an agent asked for them mid-task.
04
The privacy settings to check before you switch on an AI agent
If you try either one, a few minutes in the settings goes a long way:
On Muse: open the setup menu, go to the data controls section and switch off the option that lets Meta use your activity to improve its models. It’s on unless you change it.
On Dots: check your ChatGPT data controls. If your account allows model training, that choice carries over to your Dots.
Limit access: skip permissions the agent doesn’t need yet, like microphone or full disk access, and don’t store passwords in it. Use a separate password manager instead.
Set spending rules: on Dots, add a Custom Rule capping purchases. On either agent, read every approval prompt before tapping yes.
The next test is Meta’s promised cryptographic lock on Muse’s virtual machines, due later this year. If it arrives and holds up, Meta’s biggest privacy weakness shrinks. OpenAI, meanwhile, is expected to bring Dots to cheaper ChatGPT plans over time, which would put its safeguards in front of a much larger and less careful crowd.
Retailers are pushing back too. Amazon has blocked Muse’s shopping attempts, arguing the agent didn’t identify itself properly, and is working to block rival agents from Google and OpenAI as well.
06
Quick answers
Does Meta’s Muse train on my data?
Yes, by default. You can opt out in Muse’s data controls, but Meta can still access the data stored in your private virtual computer.
Is OpenAI Dots safer than Muse?
It has stricter approval steps and fewer reported incidents so far, but it also has far fewer users. Your ChatGPT training setting applies to it.
How much does OpenAI Dots cost?
Dots launched for ChatGPT Pro subscribers at $100 a month, with the first Dot included, plus eligible business plans.