Meta has hit pause on the program that quietly recorded its US employees’ keystrokes, mouse clicks and screen content to train its AI—but not before the whole thing leaked. An internal security notice, first reported by WIRED, warned this week that data scooped up by the program had been left sitting in the open, accessible to anyone inside the company.The timing stings: it’s been barely two months since Meta told staff there was no way to switch the tracking off.The tool goes by Model Capability Initiative, or MCI, and it landed on US employees’ work laptops in April. It logged their inputs and grabbed periodic screenshots, all to teach Meta’s AI agents how people actually navigate software—the dropdown menus, the keyboard shortcuts, the small stuff models still fumble.Opting out wasn’t an option at launch. After weeks of pushback, Meta softened slightly, letting some staff pause the tracking for 30 minutes at a stretch.
Meta’s MCI spilled data internally, revealing employee activity to all
The notice went out Monday, after a Meta engineer flagged that MCI data had spilled internally. Per WIRED, the exposure spanned roughly 45,000 hive tables holding employee activity—full prompts and transcriptions, private conversations, even people and performance records.Business Insider, which got screenshots of the incident, reported Meta logged it as a SEV 2, the second rung on its five-point severity scale.Inside the company, people were not amused. “I am incensed,” one employee wrote in an internal group, per Business Insider, angry the data hadn’t been locked down the way Meta promised. Over in a channel where staff swap jokes, someone dropped a meme: “0 days since our last nonsense.”A former employee who’d fought the program called the lapse “a mess,” and said it was exactly what workers had been flagging for weeks.
Why 1,600 Meta employees fought the AI tracking tool
MCI drew fire well before it leaked. More than 1,600 employees signed a petition warning the data collection invited security and regulatory trouble, breaches included. Zuckerberg brushed past those concerns, telling staff in leaked audio that AI learns best by “watching really smart people do things”—and that Meta’s own workforce beat any contractor it could hire.Meta insists nothing was improperly accessed. “We have carefully designed this program with privacy safeguards,” spokesperson Tracy Clayton said, “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.”CTO Andrew Bosworth conceded the rollout missed the bar its own privacy review had set.It caps a brutal stretch for Meta—mass layoffs, an AI reorg Bosworth himself called “atrocious,” and a roughly $140 billion AI bill. And it follows a rare note of contrition from the top: in a leaked memo this month, reported by Reuters, Zuckerberg told staff the company had “made mistakes and will almost certainly make more” as it pushes through its AI overhaul.
Leave a Reply