Meta paused AI training on employee keyboard input after an internal researcher misplaced sensitive keystroke logs from the Model Capability Initiative. CTO Andrew Bosworth confirmed the data was stored in an unauthorized location, prompting an immediate halt to training workflows and an internal review of data handling protocols.
Meta’s AI training operations were halted when its Model Capability Initiative (MCI) inadvertently collected and stored employee keystroke data in an unauthorized location. The pause affects ongoing experiments utilizing internal keyboard-input datasets.
An MCI researcher placed sensitive keystroke logs—intended for anonymized model testing—in a repository without proper access controls. The misplaced dataset contained unredacted input streams from multiple employees.
CTO Andrew Bosworth announced an immediate internal audit of data handling procedures and is implementing stricter repository controls. Training on the affected data has been suspended until security measures are validated and compliance protocols are reinforced.
