AI agents are now competing with promises of privacy, but Muse’s case shows that advertising claims are not the same as actual data control. Even though the data is stored in a secure VM, Meta can still access it, and the default settings still allow the data to be used for model training.
Dots, meanwhile, promotes more granular controls, such as setting rules to prevent purchases above a spending limit, along with a zero data retention option for organizations. But the lack of major privacy scandals so far does not mean Dots has been proven safe, since it remains limited to ChatGPT plans priced at $100 or more. Users should therefore examine each system’s defaults, access scope, and history of failures in full.
The Competition Between Dots, Muse, and OpenClaw
This competition is not only about AI agent capabilities. It is also about which company can make users more willing to entrust these systems with their personal data. OpenClaw came first, with Meta positioning Muse as a safer alternative. OpenAI then launched Dots with a promise that it would be safer than Muse.
The problem is that these promises still need to be proven through real-world use. Muse has faced both vulnerabilities and cases in which users did not know how far the system could access or share their data. As a result, this competition is being measured as much by transparency and data control as by features.
Allison Johnson of The Verge felt uneasy about entering banking information into a bot, even though the information was directly related to the task at hand.
This is the central challenge for AI agents: they need access to large amounts of personal data to be genuinely useful. The deeper an agent can access a user’s data, the more confident users must be that the information will not be used or shared beyond what they intended.
Where Dots Fits in OpenAI’s Product Strategy
OpenAI introduced Dots at DevDay, positioning it as an assistant that is “trustworthy, safe, and secure,” while highlighting user controls such as setting limits to prevent an agent from making purchases above a specified amount.
For enterprise customers, OpenAI offers stricter data controls, including a zero data retention option, which means data is not stored on OpenAI’s servers. This approach means Dots is selling not only agent capabilities but also confidence in using the system.
From OpenClaw to Muse, and from Muse to Dots
This progression reflects a race to raise the bar for privacy promises: Muse was positioned as a safer alternative to OpenClaw, while OpenAI presented Dots as a safer alternative to Muse. But Muse’s weaknesses show that a secure VM alone does not mean data is beyond the company’s reach.
| Factor | OpenClaw | Muse | Dots |
|---|---|---|---|
| Role in the story | The model Muse claimed was less safe | The alternative claimed to be safer than OpenClaw | The alternative OpenAI claims is safer than Muse |
| Security approach | No details in the source material | Uses a secure VM | Focuses on giving users greater control over the agent |
| Known weaknesses | No details in the source material | Meta can still access the data, and multiple vulnerabilities have been found | Its promises still need to be proven as adoption grows |
How Well Do Privacy Promises Work in Everyday Life?
Muse stores data in a secure VM, but Meta can still access it. Users therefore need to distinguish between “data is stored separately” and “the provider cannot access it.”
Muse’s default settings also allow Meta to use data for model training, even though users can opt out. In one case, the system read private messages without asking and offered a user’s address to strangers through Marketplace.
Dots, on the other hand, allows users to set rules governing the agent, such as prohibiting purchases above a specified spending limit. This approach can reduce the risks of allowing an agent to make decisions on its own, but how credible these promises are will still depend on real-world use.
| Factor | Dots | Muse | Instinct |
|---|---|---|---|
| Control | Can set rules for the agent, such as spending limits | Users do not always know the scope of its operations | Not specified in the source |
| Data | Offers zero data retention for organizations | Can use data for model training by default and build profiles of people users know | The terms of service grant very broad access to data |
| Security | No major privacy scandals have emerged so far | Has experienced a zero-day vulnerability and issues that put Meta’s internal databases at risk | No specific incidents mentioned |
| Usage conditions | Available only on ChatGPT plans priced at $100 or more | The source does not specify pricing conditions | The source does not specify pricing conditions |
| Reference point | OpenAI positions it as a safer alternative to Muse | Launched as a safer alternative to OpenClaw | Criticized over its terms of service |
Visible Strengths and Limitations That Cannot Be Ignored
Overall, both Muse and Dots have interesting privacy concepts, but their security must be proven through real-world use rather than promises alone.
Pros
- +Muse isolates data in a secure VM
- +It plans to use encryption to prevent Meta from accessing the data
- +Dots can set rules governing operations, such as prohibiting purchases above a specified spending limit
- +Organizations can choose a zero data retention policy
Cons
- −Muse has experienced multiple vulnerabilities and security issues
- −The system collects user data and creates detailed profiles of people close to them
- −The default settings allow Meta to use data for model training
- −Providing personal information to an agent still carries risks
- −Dots is included in ChatGPT plans priced at $100 or more, making its user base difficult to assess
The subscription fee is only the first cost, since Dots is available on ChatGPT plans priced at $100 or more. The real cost is the personal data users must hand over to the agent to get work done, ranging from private messages to banking information.
There is also a risk that the data could be used to train models if users do not disable the permission, or that the agent could unintentionally expose information, such as offering an address to a stranger. If a vulnerability allows an outside party to take control of the system, the impact could extend to a company’s internal data as well.
The Key Question Is Not Who Makes the Better Promise, but Who Can Actually Control the Risk
Muse, Dots, and other companies are using the word privacy to build trust, but promises alone are not enough. What matters is whether the default settings allow data to be retained or used for model training, and how clearly users can disable those options.
Users should also check what data the agent can access, whether it must request permission before taking important actions, and how the company has handled vulnerabilities or misconfigured settings in the past. Muse’s incidents show that even when a system separates data and emphasizes security, it still has to prove its claims through controls that can be independently verified.