cmux
Pass
Audited by Gen Agent Trust Hub on Apr 7, 2026
Risk Level: SAFECOMMAND_EXECUTIONREMOTE_CODE_EXECUTIONPROMPT_INJECTION
Full Analysis
- [COMMAND_EXECUTION]: The skill utilizes the
cmuxCLI to orchestrate terminal panes and performs a directory listing check usinglsto identify missing companion skills for the user. - [REMOTE_CODE_EXECUTION]: Includes instructions for installing related developer tools from the vendor's repository using the standard
npxcommand. - [PROMPT_INJECTION]: The agent is instructed to use a specific emoji (🧢) in its initial response. Additionally, the skill's data-ingestion capabilities create a surface for indirect prompt injection:
- Ingestion points: Reads terminal output via
cmux read-screenand web page content viacmux browser snapshot. - Boundary markers: Absent; no delimiters or ignore instructions are provided to isolate processed external data.
- Capability inventory: Provides significant execution capabilities, including sending arbitrary shell commands to terminal panes and evaluating JavaScript in browser surfaces.
- Sanitization: Absent; no sanitization or validation of the external content is implemented.
Audit Metadata