The Agent Skills Directory

[PRIVILEGE_ESCALATION]: The installation guide includes a command that uses sudo to create a symbolic link in the /usr/local/bin/ directory, which is a system-level modification.
Evidence: sudo ln -sf /Applications/cmux.app/Contents/Resources/bin/cmux /usr/local/bin/cmux in SKILL.md.
[INDIRECT_PROMPT_INJECTION]: The skill facilitates the ingestion of untrusted data from web pages, creating a surface for indirect prompt injection attacks.
Ingestion points: Browser snapshots, body text, and console logs retrieved via cmux browser commands as described in SKILL.md.
Boundary markers: The skill does not define or recommend the use of boundary markers or specific instructions to treat ingested data as untrusted.
Capability inventory: The skill enables full control over the terminal environment, including subprocess execution, file interaction, and browser automation via the cmux CLI.
Sanitization: There are no methods described for sanitizing external web content before it enters the agent's context.
[EXTERNAL_DOWNLOADS]: The skill instructs the user to fetch and install software from external package managers and repositories.
Evidence: brew tap manaflow-ai/cmux && brew install --cask cmux and npx skills add manaflow-ai/cmux in SKILL.md.
[COMMAND_EXECUTION]: The skill's primary functionality is based on executing shell commands and interacting with a local JSON-RPC API via a Unix socket.

cmux