Bosworth's first on record explanation of Meta's Model Capability Initiative pause names one researcher, not a breach. The keystroke to AI pipeline that fed it is what actually matters.
Meta paused an internal AI training program that captured every keystroke and mouse movement from most US employees after one researcher's mishandling pushed the data into an internal location where it did not belong. That is the on-record account from Meta.
"No breach," one researcher, the program "quite secure" and limited to a small number of people. That is the framing Bosworth laid out in the Atlantic interview. On those terms, the Model Capability Initiative was a small data-governance failure. On the program's own design, the failure mode is structural: a surveillance pipeline whose intended output is model training data has "internally" as the only place its data can leak to.
The keyboard and mouse logger captured keystrokes and cursor activity to feed Meta's AI models. TechCrunch framed the launch as a training-data play. Wired subsequently reported Meta accidentally exposing employee keystroke data across the organization, surfacing a cross-employee access pattern the rollout did not acknowledge.
Bosworth's account does not contradict either report. He told Thompson the captured data, in a transformed state, "landed someplace that it shouldn't have landed internally," and that Meta responded by locking the program down pending review. The "someplace" he described is the surface where the pause actually took place: inside Meta's own data-handling boundary, not outside it. The internal destination, not an external one, is the failure the program was paused for.
That detail is easy to lose in Bosworth's preferred frame. He used "no breach" twice in the interview, said only one researcher was involved, and called the program "quite secure" by design. A reader working only from the executive quote would reasonably conclude the program had a contained incident and the perimeter held. The mechanism that made an internal exposure possible does not feature in Bosworth's account.
A keystroke-and-mouse logger is, by definition, a recorder of what an employee types and where they click. If the resulting stream is accessible to more than a small authorized group, a one-off mishandling is the program's normal operating surface rather than an exception. Wired found exactly that kind of access. Bosworth's "small number of people" with access is consistent with a permeable-by-default architecture even as it minimizes the scope of the incident.
The pause itself was already in motion. The BBC reported Meta scaling the program back amid internal pushback, dialing down capture rather than fixing a specific mishandling. Bosworth's interview is therefore less a breaking story than a cause-and-effect account. It does not name what happens next.
Three operational questions the interview does not answer: whether the Model Capability Initiative is paused, canceled, or being redesigned in place; whether employees will receive any review or audit access to what was captured about them during the rollout; and whether the training-data set built before the pause is being kept, deleted, or rebuilt under a different governance framework.
For a company Wired separately described as posting record profits alongside record-low morale in a year of AI-linked reorganizations and layoffs, those are the answers that will determine whether "Model Capability Initiative" reads, in a year's time, as the name of a paused experiment or as the name of a pipeline that briefly captured what every US Meta employee typed.