Researcher discovered the internal structure of the AI-agent Meta Muse

An independent security researcher gained access to the 6.8 GB Meta Muse file system which was mistakenly transferred during interaction with the system. Inside the archive were full records of 113 sub-agents and 68 specialized skills used by the model to perform tasks. Analysis of the dump showed the presence of configurations for integration with external services, including Slack and Dropbox, which had not been officially announced before. The documentation reveals internal protocols for agent interaction and context management methods that Meta applies to automate workflows within its ecosystem. The file system also contains detailed logs of command execution and data structures used to orchestrate agent cycles. The researcher noted that the data includes not only technical parameters but also specific instructions for processing corporate communications, which gives an idea of how the company plans to scale autonomous agents. This incident allows us to evaluate the real architecture of Meta's agent systems, which until now remained closed to external analysis. Details of the technical analysis of the Muse structure have already sparked discussions in the developer community regarding the security of storing configuration files in cloud AI tools.