Overview
- Meta began pausing the Model Capability Initiative after an internal security notice revealed that databases containing employee activity were accessible more widely inside the company.
- The tool collected detailed inputs from U.S. staff — mouse movements, click locations, keystroke timing and occasional screen captures — to produce training data for AI models.
- Internal documents reviewed by reporters showed the exposure touched roughly 45,000 hive tables and included prompts, transcriptions, private conversations and people and performance data.
- Company spokespeople say there is no indication so far that employees improperly accessed the exposed files and engineers applied a patch after a vulnerability was first found on June 18, but the pause is being rolled out gradually.
- The incident sharpened existing employee objections, recalled a petition of more than 1,600 signatories, and raises questions about Meta’s data governance, workplace consent and possible regulatory scrutiny if the program is resumed.