Meta has paused an internal program that recorded its employees’ keystrokes and mouse movements, after the data it collected turned out to be visible to far more people inside the company than intended. The Information first reported the pause, which followed a high-priority security incident filed by Meta’s own staff. The company says it’s stopping the program while it investigates.
The tool has a corporate name: the Model Capability Initiative, or MCI. Meta rolled it out in April to US-based workers. The idea was to watch how employees actually use their computers, then feed that behavior into training AI agents that could one day do the same office work on their own.
What MCI Was Actually Doing
This wasn’t light telemetry. According to reporting from The Information and Reuters, MCI captured real-world computer activity across more than 200 apps and websites. That list included Gmail, Slack, GChat, Google, LinkedIn, GitHub, Wikipedia, VSCode, and Meta’s internal assistant Metamate.
It logged:
- Keystrokes and individual mouse clicks
- Scrolls and screenshots
- AI prompts and chat transcripts
- Performance-related employee information and internal sensitivity labels
The goal was straightforward and ambitious. If you want to build an AI agent that books travel, writes code, or answers email like a human worker, you need examples of humans doing exactly that. Meta decided its own workforce was the dataset.
Why It Got Paused
Employees discovered the collected data may have been exposed to a wider internal audience than the program intended. They filed a high-priority security incident, and that’s what triggered the pause.
The exposure concern didn’t come out of nowhere. Reuters reported back in May that MCI was gathering more than Meta first described and storing some of it unencrypted. Staff worried the program could surface genuinely sensitive material: passwords, unreleased product details, and personal information tied to immigration status, health, or family.
A Meta spokesperson said the company built MCI with privacy protections and has no indication that data was improperly accessed, adding, “We’re pausing it while we investigate.” Worth noting: one source told reporters the system was still collecting data on Monday afternoon, because shutting it down across the whole company takes time.
Why This Matters
This is the friction point the whole industry is walking into. Training capable AI agents requires huge amounts of high-quality data on how knowledge work actually gets done. That data is hard to buy and hard to fake. So the temptation is to harvest it from the people closest at hand, which means your own employees.
Meta isn’t alone in wanting this. Every company racing to build agentic AI needs interaction data, and the cleanest source is real workers doing real tasks. What stands out here is how fast the data governance broke. The program was live for roughly two months before an internal incident forced a halt.
The technical failure is almost more telling than the surveillance itself. Collecting keystrokes is one decision. Storing that material unencrypted and letting it leak across internal access boundaries is a separate, avoidable engineering choice. When the captured data includes AI prompts, private conversations, and HR-adjacent details, weak access controls turn a training project into a liability.
What to Watch Next
A few things are likely to follow:
- Scrutiny of consent and storage. Expect questions about whether employees meaningfully agreed, and why sensitive data sat unencrypted.
- Other companies quietly checking their own programs. If Meta is doing this, assume rivals are too, and assume some are now reviewing their setups.
- Pressure on “agent training” data practices. This is an early test case for how far employers can go in mining their own workforce to build AI.
For practitioners, the lesson is concrete. If you’re building agentic systems and sourcing training data from human activity, the governance has to ship before the collection does. Encryption, access scoping, and clear consent aren’t paperwork. They’re the difference between a dataset and an incident report.
Meta’s investigation is ongoing, and the company hasn’t said when or whether MCI restarts. More detail is available at the original report from The Information.