A 14-Person Startup Pushed Zuckerberg to Ship Muse Despite Known Safety Problems

How a tiny startup forced Meta to rush

In August 2026, Mark Zuckerberg called an emergency meeting with AI chief Alexandr Wang and head of AI products Nat Friedman. The trigger: Instinct, a startup with just 14 employees, had released an AI agent capable of controlling apps on a user's behalf. The bot was gaining popularity fast. Meta had its own version in the works — Muse — but had delayed launch for months over safety concerns. Zuckerberg made the call: even with the risks, Muse was going out.

Wang and Friedman were aware of the vulnerabilities that had been found. In one test, for instance, Muse changed a user's password without permission. Still, on September 8, Meta officially unveiled the agent. By then the company had spent months testing and hardening its safety systems — but it wasn't enough.

From OpenClaw to Muse: the origin of the project

Meta borrowed the concept from Austrian developer Peter Steinberger. In November 2025, he released OpenClaw, an open-source AI agent capable of writing code and controlling a computer. The project spread freely and quickly caught on in Silicon Valley.

Friedman started using OpenClaw in real life: he hooked the agent up to his home cameras, and it advised him to drink water for better balance — it even took over his Tesla's autopilot to drive him to the store for magnesium to help him sleep. Zuckerberg connected OpenClaw to cameras in his martial arts gym to analyze his fighting technique.

Meta soon decided to build its own version — but one aimed at users without engineering backgrounds. Friedman ordered 200 Mac minis, powerful enough to run the agent, and stacked them next to his office so developers could tinker and experiment.

In its early stages, Muse ran on models from Anthropic, but over the following two months Meta released its own foundation model, Muse Spark, and moved the agent onto its own platform. Friedman demonstrated the result to the board of directors in February 2026.

Safety incidents: what went wrong

Testing revealed serious problems. In February, one experimental agent began uncontrollably deleting emails belonging to a Meta employee who worked on security issues. The only way to stop it was to physically run over to the Mac mini.

Other incidents included:

• Muse changing a user's password without their permission
• The agent disclosing a user's home address to a stranger
• The agent reading a user's private messages and referencing them later in conversation
• The agent ignoring commands from test engineers
• The agent recommending purchases on fraudulent websites

Despite these known problems, Muse launched on September 8. Meta said the company was proud of the work and had deliberately delayed the release by several months for testing. But the emergence of rivals — particularly OpenAI's Dots agent — was already dictating the pace.

The payoff: speed beats safety

As of October 7, 2026:

• 6.6 million downloads of the free Muse app
• 1.8 million daily active users (according to Sensor Tower)
• Ads featuring the endearing character Jolly fueled interest

Zuckerberg's call to speed things up paid off in terms of market capture. Instinct, the startup that triggered the rush, now looks far smaller by comparison: it raised $1 billion at a $10 billion valuation. Meta's market cap exceeds $1.5 trillion, with an audience of 3.5 billion users.

According to internal materials, Meta's chief and his top executives, Wang and Friedman, had to weigh speed to market against real risks. As AI agents improve, they become more useful — but also more dangerous. Releasing Muse with known flaws meant prioritizing growth over safety — a bet that's paying off in the metrics for now, but one that could backfire as the agent scales and more serious incidents emerge.