Meta pauses keystroke-tracking AI training program after internal leak

Meta is pausing its Model Capability Initiative (MCI), an internal program announced in April that uses staff keystrokes and mouse movements as AI training data, after sensitive data was found accessible across the entire company. Screenshots obtained by Business Insider showed the leak exposed employees' private conversations, performance data, and transcriptions. Meta classified it a SEV 2 on a 0-to-5 scale (0 being most severe) and confirmed it is investigating.

The program was already controversial: mandatory for most staff, it drew backlash when announced because employees were uncomfortable having their activity recorded. The leak inflamed that anger — one employee wrote 'I am incensed' in an internal group, with colleagues criticizing that the data wasn't locked down from the start. A Meta spokesperson said it 'carefully designed this program with privacy safeguards' and has 'no indication at this time that any data was improperly accessed,' but is pausing pending investigation.

The incident is the latest in a string of Meta security problems, following a flaw that let people hijack Instagram accounts and a rogue AI agent that caused a severe incident in March. It also sharpens the irony of Meta harvesting employee behavior to train models meant to automate knowledge work, even as its 'AI for Work' product head departed this week. Watch whether MCI resumes with new controls or is scrapped, and whether the leak triggers regulatory interest in employee-surveillance-as-training-data.