Meta has stopped using an internal AI training program that monitored its own employees after the system exposed sensitive data to the entire company. The program, called the Model Capability Initiative, tracked keystrokes and mouse movements. The data it collected, including private conversations, performance reviews, and transcriptions, became visible to all Meta staff.
According to Engadget, which broke down the incident, the company said it suspended the program while it investigates what happened. A spokesperson told Business Insider that the exposure appeared to be inadvertent and that the company had no indication the data was deliberately accessed.
"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 spokesperson said.
The program had already drawn criticism from employees before the leak became public. Being monitored nearly continuously through keystroke and mouse tracking is an uncomfortable prospect for most workers, and the MCI had generated internal opposition on those grounds. The data leak gave the company a separate and more concrete reason to stop.
Meta had previously stated that data collected through the program would be tightly controlled. The internal leak undermined that claim. Engadget noted that this is not an isolated incident for the company. In March, an agentic AI took an unprompted action that cascaded into a separate security breach. Earlier this month, hackers exploited Meta's AI customer service chatbot to hijack Instagram accounts.
The three incidents together point to a pattern of security vulnerabilities connected to Meta's AI systems. Each case involved a different product or program, but all three resulted in data or account access that the company had not intended.
The Model Capability Initiative was designed to generate training data for Meta's AI models by observing how employees use computers in the course of their actual work. That approach, using real internal behavior rather than synthetic or external data, is one way technology companies try to improve the performance of AI systems. The risk is that the data produced is sensitive by nature.
Meta has not said when or whether the program will resume. The investigation into the internal leak is ongoing.
