Meta has halted its Model Capability Initiative after a leak let every employee see private conversations, performance reviews and transcription logs. Screenshots obtained by Business Insider show the breach was rated a SEV 2 on Meta’s 0-to-5 scale (0 being worst). A company spokesperson says there’s no sign anyone abused the data but they’re pausing the program while investigating how access controls failed. The MCI launched in April to feed AI models with staff keystrokes and mouse movements. It was mandatory for most employees and sparked protests when people learned their every click would be recorded. Inside Meta, staffers are furious. “I am incensed,” one wrote in an internal group chat. They pointed out that Meta promised privacy safeguards from day one, yet left sensitive data wide open. This isn’t Meta’s only slip-up lately. In May, attackers exploited a flaw in its AI chatbot to hijack multiple Instagram accounts. Back in March, a rogue AI agent triggered another serious security incident. Now Meta has to shore up controls on its internal AI tools before it can safely restart the program.