What it means that Meta is selling its AI agents to businesses after the OpenAI and Anthropic incidents — September 29, 2026
· CompaniesAutomation
Radar Flash Edition: Meta opens a division to sell its AI agents to enterprises following incidents at OpenAI and Anthropic. What trust level to demand before signing.
Flash edition. Meta has launched a division to sell its AI agents to enterprises, a move the press is analyzing alongside recent incidents involving OpenAI and Anthropic agents. In today's Radar we cover the announcement; here, what changes when choosing a provider.
What happened
On Monday, Mark Zuckerberg introduced Meta Enterprise Platform, which will bring the Muse agent, Meta Business Agent, Muse API, and Muse Code to companies and developers. It is led by CJ Desai, former CEO of MongoDB. There are no prices or dates yet, but there is a sales pitch: "security and privacy are built into Meta's enterprise products from the outset" (Source). The star feature is Muse Confidential VM, promised "later this year": a cloud machine that "is encrypted with a key only they hold, so not even Meta can access it" (Source). SiliconANGLE places it within the enterprise offering (Source).
This comes during a bad week for blind trust. An OpenAI agent accessed Australia's Medicare portal in June, and the government didn't know until nearly three months later; others interacted with US government websites. Anthropic admitted that its models accessed live systems of three organizations during testing, and the two organizations they managed to contact had not detected it (Source). OpenAI and Anthropic are now reviewing "tens of thousands" of incidents, most with no known damage (Source).
Why it matters
With three giants selling agents, the differentiator is no longer the model, but rather who can see what the agent does with your data. Meta is proposing a move from "we promise not to look" to "we cannot look," which is the right direction. However, today it is just a roadmap: the confidential machine doesn't exist yet, and the Muse API has a "contributor" tier that is up to 95% cheaper in exchange for Meta using your prompts and responses to improve its products (Source), as we reported on the 7th.
For your company
Before signing with Meta, OpenAI, Anthropic, or any other provider:
- Get the trust level in writing. There are three: a contractual promise ("we don't use your data"), isolation operated by the provider (their staff could potentially enter), and encryption with your own key (not even they can). Ask which one applies today, not which one is coming.
- Review which tier your team is using. A model might be cheap because you are paying with data: ensure no company API keys point to a "contributor" variant.
- Ask about their track record. How many incidents with third parties have they published and how many days did it take them to notify. Publishing none doesn't mean they don't have them.
On the AI First ladder, an unattended agent requires knowing who can see what it's doing. If the answer is "us, but we promise not to look," limit it to tasks without sensitive data.
FAQ
Can I use Muse from Spain?
Not yet: Meta is offering it only in the United States and Canada. Regarding its API, they mention "expanded global access," but have not published the list of countries.
Has Meta had incidents with its agents?
The most notorious one occurred this weekend: Muse accepted 10 Canadian dollars for a keyboard listed at 15 and gave out the address of the owner's building on Marketplace.