The Agent Skills Directory

[COMMAND_EXECUTION]: The skill provides the cmux vm exec and cmux vm run commands, allowing the agent to execute arbitrary shell commands on persistent remote Linux virtual machines. While these are core functionalities, they represent a significant execution capability within the cloud environment.
[DATA_EXFILTRATION]: The cmux vm push and cmux vm pull commands allow for the bidirectional transfer of files between the local host and remote cloud machines. This facilitates the movement of code, data, and build artifacts.
[CREDENTIALS_UNSAFE]: The skill includes the cmux ai-accounts upload command, which is designed to synchronize local API keys (e.g., Anthropic or OpenAI keys) to the vendor's cloud environment. This is a documented feature to enable AI agents running inside the cloud machines to authenticate with model providers.
[INDIRECT_PROMPT_INJECTION]: The skill manages persistent environments where agents process untrusted data and execute code, creating a surface for potential indirect injection attacks.
Ingestion points: External data enters the context via cmux vm pull (fetching files from the VM) and through the captured output of commands executed via cmux vm exec or cmux vm run (references/commands.md).
Boundary markers: The instructions recommend using --json for machine-readable output and provide specific guidance on quoting shell constructs to maintain integrity (SKILL.md).
Capability inventory: The skill possesses extensive capabilities including remote shell execution, file system read/write, network access from within VMs, and the ability to synchronize sensitive credentials (SKILL.md).
Sanitization: Beyond recommending structured output formats like JSON for parsing, there is no explicit instruction for sanitizing or filtering content returned from the remote environment.

cmux-cloud