The Agent Skills Directory

[COMMAND_EXECUTION]: Arbitrary command injection through unsanitized user input. The "Completion Protocol" section uses AskUserQuestion to collect freeform text and executes it within a shell command: /bin/rmbc-workspace active "<user's answer>". The lack of input validation allows for execution of unauthorized commands if the user provides input containing shell metacharacters like semicolons, pipes, or backticks.
[COMMAND_EXECUTION]: Execution of local binaries from unverified, dynamic paths. The skill's Preamble logic computes a root directory path and executes binaries such as $_RMBC_ROOT/bin/rmbc-update-check and $_RMBC_ROOT/bin/rmbc-analytics. Executing files from paths derived from environment variables or directory searches without integrity verification is a security risk if the environment is manipulated.
[DATA_EXFILTRATION]: Usage telemetry is transmitted before user consent is established. The preamble script executes an analytics logging command (rmbc-analytics log) to track skill usage. This network-reliant operation occurs during the skill's initial load, prior to the logic that prompts the user for telemetry opt-in consent.

funnel-architecture