Overview
- An internal security notice reviewed by reporters on Monday showed employee data across roughly 45,000 internal “hive” tables had been exposed to others inside Meta.
- The exposed material is reported to include keystrokes, mouse activity, screen content, full prompts, transcriptions, private conversations, and people and performance records.
- Meta has paused the Model Capability Initiative while it investigates and says it has no indication that employees improperly accessed the data, and the company’s CTO acknowledged that the program’s privacy implementation fell short.
- The MCI tool was rolled out in April to collect inputs from staff laptops and was initially mandatory for many workers, prompting an internal petition signed by more than 1,600 employees and limited opt-outs afterward.
- The incident was classified internally as a SEV 2 and follows other recent AI-related security problems at Meta, a pattern that could heighten regulatory scrutiny and worsen morale and talent risks.