According to news on June 23, Meta has suspended an internal AI training project indefinitely. The project records keyboard input, mouse operations, and screen content on employees’ work computers to train AI to learn to operate the software; the reason for the suspension was that access rights to the project data had been misconfigured, and about 45,000 data tables were once open to Meta internal personnel.
WIRED and Business Insider separately reported on the internal permissions incident. According to Meta’s internal security notice, the data in question belongs to the Model Capability Initiative (MCI), which contains traces of operations on employees’ work computers, as well as sensitive content such as “complete AI prompts and transcripts, private conversations, personnel and performance data.” The data mentioned in WIRED's report mainly involves company equipment belonging to Meta's US employees.
Meta spokesperson Tracy Clayton confirmed that the company is investigating the security incident and said MCI has been suspended indefinitely while the investigation proceeds. She also emphasized that Meta "currently has no indication that the data has been improperly accessed by employees." However, this is only a statement at the current stage of investigation and does not equal the final conclusion.
The risks that 1,600 employees were worried about turned out to be true.
MCI was launched in April this year with the goal of collecting employee work behavior data on computers for training AI systems. Meta CEO Mark Zuckerberg previously explained this idea in internal meetings: AI models can learn by observing how good employees work, and Meta employees are better suited to provide high-quality samples than outsourced personnel hired specifically to produce this kind of data.
This explanation did not convince most employees. In May, more than 1,600 Meta employees signed an internal petition calling for an end to this computer behavior tracking. The petition states that collecting such data poses security and regulatory risks, including data breaches and unauthorized disclosure.
Meta Chief Technology Officer Andrew Bosworth previously told employees that MCI data management is "strictly controlled," using the same protection standards, storage systems and access controls as the company's other sensitive data sets. But after this incident, Bosworth admitted that the actual implementation of the project did not meet the standards listed in the privacy review document. He said in an internal post that the problem lay in a misconfigured access control list and that the company needed to find out the cause and trace the data access path.
It's not an external attack, but internal permissions are not locked properly.
This incident was classified as SEV2 internally by Meta. According to the explanation in the report, Meta's internal severity level ranges from 0 to 5, with 0 being the most severe and SEV 2 being a higher level. Internal posts seen by WIRED show that the incident has been marked "closed," meaning the permissions issue may have been fixed, but Meta has not yet publicly stated the specific fixes and full timeline.
For employees, the point of dispute is not just whether anyone has actually viewed other people's data, but rather that the company has previously promised to "strictly control" sensitive data sets, and indeed there have been issues with access permissions that are too wide. One employee wrote on an internal forum that he could see no evidence of malicious access, but it was "angry" that the data was not locked as promised. BusinessInsider also reported that internal employees reacted strongly to the project and this incident.
The incident also touches on longer-term regulatory constraints facing Meta. Meta remains subject to a U.S. Federal Trade Commission (FTC) consent order that requires the company to maintain processes to prevent data breaches until 2040. WIRED also mentioned that Meta has begun to involve AI in some privacy and security review work; however, there is currently no evidence that AI was involved in this MCI permission configuration error.
AI sprints are hitting a cost of trust within companies
MCI is not an isolated project. Meta created a new "Applied AI" (Applied AI) team in March this year, assigning approximately 6,500 employees to relevant positions, with the goal of improving AI models. Business Insider reported that some employees felt that the projects they were assigned were trivial and depleted morale; Bosworth apologized to employees last week for communication issues during the AI reorganization and promised to improve communication and restore some office benefits.
This time MCI is suspended, the internal cost of Meta's sprint towards AI is more concretely exposed. The company can say that it has found no signs of improper access so far, but as long as this kind of data has been mistakenly opened, employees' trust in "training AI with their own work behaviors" will be difficult to repair with internal promises alone. Employees are not only the providers of AI training data, but also the direct bearers of privacy and security risks. This dual identity is what really stings about this incident.