Meta pauses AI training program after employee keystroke data was left open to the whole company [[[PAUSE] NOW!!!]]

Meta pauses AI training program after employee keystroke data was left open to the whole company.

An internal notice seen by Wired said data from Meta's Model Capability Initiative, including keystrokes, screen contents, and AI prompts, was accessible across 45,000 internal tables, traced to misconfigured access controls.

Meta is indefinitely pausing a program that tracked how its US employees use their work laptops, after an internal security notice found the data it collected was accessible to anyone inside the company. The effort, known as the Model Capability Initiative (MCI), logged keystrokes, mouse clicks, and on-screen content to train AI systems to use software the way people do. According to Wired, which reviewed the internal notice and spoke with current employees, the exposed data spanned 45,000 internal "hive" tables and included full AI prompts and transcriptions, private conversations, and people and performance data.

A Meta spokesperson confirmed the company is investigating and said it is pausing the collection indefinitely, adding that there is no indication so far that any employee improperly accessed the data. Internally, the explanation was more specific. Chief technology officer Andrew Bosworth told staff the rollout had fallen short of the program's own privacy review and pointed to misconfigured access control lists as the cause, saying the company needs to trace every data access and understand how it happened.

That admission is awkward against the record. Only a couple of months earlier, according to internal posts seen by Wired, Bosworth had assured concerned employees that the tracking was "tightly controlled" and used the same protections and access controls as Meta's other sensitive datasets. Employees noticed the gap, with one forum post joking about the company's running streak of self-inflicted problems and others asking how the privacy reviews had failed.

How it got here

The program began in April and drew resistance almost immediately. Last month, more than 1,600 employees signed an internal petition protesting the laptop surveillance, warning that the collection created both security and regulatory exposure and could lead to a breach or improper disclosure. One engineer wrote a widely shared note describing the screen-scraping as an invasion of privacy.

Meta executives defended the project as necessary to train AI on real human computer use. In leaked audio of a company meeting last month, chief executive Mark Zuckerberg argued that AI learns by watching skilled people work, and that Meta's own staff were a higher-caliber source than contractors the company could hire to generate the same data. After the backlash, Meta began offering limited exemptions this month, including letting staff briefly switch off the monitoring for sensitive personal tasks. Some employees are still pushing for it to be scrapped entirely.