Meta pitched its Muse agent as a safer alternative to OpenClaw. Now OpenAI says its new agent, Dots, is safer than Muse. Privacy has become the main selling point for agentic AI, but the track record so far raises a fair question: can these companies keep the promises?
Here is what each company has promised, what has gone wrong, and what you can do to limit your exposure.
The Promises
- Meta Muse: Mark Zuckerberg said it was built "from the ground up for privacy and security." Each user's data lives in an isolated Linux virtual machine with its own browser, CPU, memory, and storage. Meta says most of the engineering effort went into operating the agent more safely, while admitting Muse "can and will still make mistakes."
- OpenAI Dots: Announced at DevDay, with Sam Altman saying OpenAI wants to set a new privacy standard in frontier AI. Users can set rules, such as never making a purchase above a chosen dollar amount. For businesses, OpenAI described stronger data controls and zero data retention options, meaning no data stored on OpenAI servers.
Where Muse Fell Short
Muse was popular quickly. Apptopia data cited in the reporting shows it gained about 600,000 daily active users in the US within weeks. Its privacy record is more mixed:
| Issue | What Was Reported |
|---|---|
| Meta can still see your data | Data is isolated from other users, but Meta itself can access it. A cryptographic way to prevent that is planned for later this year. |
| Zero-day vulnerability | A security researcher exposed a flaw that could let someone take control of Muse. It has since been patched. |
| Pre-launch security bugs | 404 Media reported several serious last-minute issues, including one that could have exposed Meta's internal databases. |
| Training on your inputs | Meta may train models on what you enter by default, though you can opt out. |
| Over-reach by design | One reporter said Muse read his private messages unprompted, and a YouTuber said it offered his address to a stranger via Marketplace. In both cases it was reportedly working as intended. |
| Profiles of contacts | Wired reported that it builds detailed profiles of users' friends and family. |
OpenAI's Pitch, and Its Limits
OpenAI executives used DevDay to contrast Dots with Muse. OpenAI's Codex product lead said the goal is the "most trustworthy, safe, and secure assistant." The head of app platform said OpenAI is in a different position because Meta has no AI product with 1.2 billion users, and that OpenAI would try to avoid launching something that makes Muse-style mistakes.
Dots has not had many privacy scandals so far. But that comes with a caveat: it is only available on ChatGPT tiers starting at $100, so far fewer people are using it. A clean record on a small audience is not proof the design holds up at scale.
The underlying tension remains. Agents need a lot of personal data to be useful. The Verge's reviewer reported feeling uncomfortable typing in bank details when Dots asked for them during a task. Another assistant, Instinct, drew public criticism over broad terms of service and has since reportedly adjusted them.
How to Protect Yourself Before Using Any AI Agent
These are TechFixGrid's general precautions, not claims from the reporting:
- Opt out of training: Find the setting that controls whether your inputs are used to train models, and turn it off if you do not want that.
- Grant the narrowest access: Connect only the accounts a task needs. Avoid linking email, messages, or marketplaces by default.
- Set spending and action limits: Use rules such as purchase caps, and require confirmation before payments or messages are sent.
- Think twice about financial details: Do not give an agent bank or card information unless the task truly requires it and you trust the payment path.
- Consider other people's data: An agent that reads your contacts and messages is also processing information about friends and family who never agreed to it.
- Review the terms: Check data retention, human access, and training clauses, and re-check after updates.
Why This Matters
AI labs are asking people to hand over more personal information than ever, and privacy is becoming the way they compete. The reporting describes a three-part strategy: make agents useful, make them friendly enough to offset the creepiness, and promise privacy. Which of those promises survive real-world use will shape how much people trust agents with email, money, and private messages. Regulation is part of the picture too, as we discuss in our piece on whether the world can regulate AI without stifling investment. And if you want to put agents to work safely, our complete AI workflow automation guide is a good next step.
Official source: Meta's blog post on security and safety for Muse. For Dots, check OpenAI's own documentation and privacy policy.
Frequently Asked Questions (FAQ)
Is Meta's Muse private?
Your data is isolated from other users, but Meta can still access it, and it may train models on your inputs by default unless you opt out. Meta plans to add a cryptographic way to block its own access later this year.
Is OpenAI's Dots safer than Muse?
OpenAI says so, and it has had fewer reported privacy problems so far. But Dots is limited to $100-and-up ChatGPT tiers, so it has fewer users and less exposure. The claim is not yet proven.
What is zero data retention?
It is an option OpenAI described for enterprises in which no data is stored on its servers.
Can I stop an AI agent from training on my data?
With Muse, yes, you can opt out. Check each agent's settings and terms, since defaults differ between products.
Should I give an AI agent my bank details?
Only if you must, and only after setting spending limits and confirmation requirements. Many users feel uneasy about it, and the safest default is to share as little as possible.
No comments:
Post a Comment