Overview
- Meta paused the Model Capability Initiative after a high‑priority SEV2 report on Monday showed employee data was exposed internally across thousands of tables.
- The MCI tool, rolled out in April, records mouse movements, clicks, keystrokes and occasional screen content from U.S. employees to train AI models and was initially mandatory for most staff.
- Internal documents and security notices say the exposure touched roughly 45,000 hive tables and included prompts, transcriptions, private conversations and personnel records.
- CTO Andrew Bosworth acknowledged the rollout fell short of privacy‑review standards, and Meta says there is no current indication of improper access while the company investigates and phases the pause.
- The incident has intensified employee anger after more than 1,600 staffers protested the program, and it could prompt regulatory review and changes to how Meta balances rapid AI training with data safeguards.