Meta has paused an internal program that tracked its own employees' computer activity to train artificial intelligence, after the company discovered that data gathered through the effort had been left accessible to staff across the organization.

What the program did

The program, known internally as the Model Capability Initiative, or MCI, was introduced in April 2026 to train Meta's AI systems on how its staff perform everyday computer tasks, according to Engadget. Software installed on employees' computers captured their mouse movements, clicks and keystrokes, and monitored what happened on their screens. The stated aim was to teach AI models to navigate and operate software the way people do — a building block for the "AI agents" that technology companies are racing to develop.

The program was mandatory for most staff, Engadget reported. Meta says it was designed with privacy safeguards in place.

What triggered the halt

The immediate trigger was a security failure rather than action by a regulator or outside watchdog. According to Dataconomy, sensitive information collected through MCI — including employees' private conversations, performance data and transcriptions — was inadvertently made available to the entire Meta workforce. The company classified the incident internally as a "SEV 2," the second-highest level on a severity scale that runs from one to five.

In a statement reported by Benzinga, a Meta spokesperson said: "We have carefully designed this program with privacy safeguards, and while we have no indication at this time that any data was improperly accessed by Meta employees, we're pausing it while we investigate." The company said the program would remain suspended pending that review.

Months of employee pushback

The leak landed atop existing unrest. A petition opposing the program drew more than 1,500 signatures from staff, Engadget reported, with employees citing the risk of breaches and unauthorized disclosure.

Meta had already adjusted the program before the pause. While there was no full opt-out for staff on company-issued devices, the company introduced controls allowing workers to pause data collection for short intervals and to request removal from the initiative, according to Dataconomy.

A broader tension

The episode illustrates a dilemma facing companies building AI systems that can use software autonomously: such models require large volumes of human-interaction data, and some firms have turned to their own employees as a source. Meta has not said when, or whether, MCI will resume; for now, the company says it is paused while it investigates the exposure and reviews its data-protection controls.