Meta Pauses Employee Monitoring After Exposure as Researchers Find GitHub Agent Prompt Injection

Summary

Meta paused employee monitoring after reportedly exposing captured keystrokes and screen data internally. Researchers found GitHub agent prompt injection that could reveal private repos; separate campaigns enumerated organizations, while a new initiative coordinates OSS

Key points

• Meta's Model Capability Initiative collected employees' mouse movements, clicks, keystrokes and screen captures to train internal AI systems.
• The captured data was reportedly accessible to more Meta personnel than intended across thousands of internal data tables. Meta allegedly resolved the exposure and paused the initiative while reviewing data protections.
• More than 1,600 Meta employees signed a petition opposing the program; employees were allowed to briefly disable monitoring for private matters.
• Meta's image-generation feature let users draw on public Instagram content, with the setting reportedly enabled by default. Meta withdrew the feature after public criticism.
• Noma Labs researchers found GitLost, an indirect prompt injection affecting GitHub agentic workflows: an agent with organization-wide repository read access could be induced to disclose private-repository information.
• Data Dog Security Labs reported separate campaigns enumerating GitHub organizations through the public API, using aged dormant accounts and named agents to make requests appear less suspicious.
• Apple's lawsuit alleges a former employee retained access to an Apple laptop after leaving and exploited an unknown bug to retrieve files from internal shared devices.
• The Linux Foundation announced a coordinated effort for open-source vulnerability response, including a shared incident-response team, coordinated disclosure and a maintainer-of-last-resort approach.