Meta has paused an internal AI training program after sensitive employee data from the project became broadly accessible inside the company, according to Business Insider reporting based on screenshots it obtained.
The program is called the Model Capability Initiative, or MCI. It was designed to help Meta train AI models on how people actually use computers by collecting signals such as keystrokes, clicks and mouse movements from employees' work machines.
What leaked
Business Insider reported that a screenshot showed the exposed data included employees' private conversations, performance data and transcriptions. Meta classified the incident as SEV 2 on its internal scale, where SEV 0 is the most severe.
Meta confirmed the incident and said it is investigating. The company's public line is that it has no indication, so far, that the data was improperly accessed by employees.
Why employees were already uneasy
MCI was controversial before the exposure. The program was mandatory for most staff, and employees had already raised concerns about turning their everyday work behavior into AI training data.
The core concern is not just that Meta recorded interaction data. It is that work laptops can contain messy, mixed-context material: internal chats, prompts, transcripts, personnel information, personal records accessed from a work device, and performance-related discussions. Once that material is collected for model training, the permission boundary has to be unusually tight.
That is where this incident landed. The pause was triggered by a data-access failure, not by a philosophical reversal on employee monitoring.
What Meta says now
Meta says it carefully designed the program with privacy safeguards and is pausing it while the investigation runs. NDTV, summarizing the Business Insider report, said Meta positioned the pause as a temporary step while it checks what happened.
LiveMint's follow-up coverage, citing Reuters-reviewed documents, added that the exposed material included full prompts and transcriptions, private conversations, people and performance data, and sensitivity ratings. It also reported that stopping the tool across all employees would take time to roll out.
Why this matters
The story is a clean warning for every company building workplace AI agents: training data gathered from employee desktops is not ordinary telemetry. It is behavior, context, intent and sometimes sensitive content bundled together.
If that data is used to teach AI agents how to operate software, companies need more than broad consent language. They need strict access controls, aggressive filtering, retention limits, audit logs and clear separation from performance management. Without that, AI productivity programs can quickly look like workplace surveillance with a model-training label.
What to watch next
The key question is whether Meta restarts MCI with narrower collection and stronger access controls, or whether the backlash makes large-scale employee activity capture too expensive internally.
For AI teams, the practical lesson is immediate: if your agent-training plan needs real employee activity data, design the privacy controls before the rollout, not after a SEV report.