The Agent Skills Directory

[COMMAND_EXECUTION]: The skill enables the agent to execute shell commands in terminal panels and JavaScript within browser surfaces. These are core features of the tool designed for managing development servers and automating browser interactions.
[DATA_EXFILTRATION]: The skill includes patterns for reading browser-resident data, such as localStorage items, as part of its automation and state-checking features.
[PROMPT_INJECTION]: The skill presents an attack surface for indirect prompt injection through its data ingestion capabilities. (1) Ingestion points: The agent reads external data from terminal outputs using cmux read-screen and from web content via cmux browser snapshot or cmux browser get. (2) Boundary markers: The provided usage patterns do not include explicit delimiters or instructions to treat ingested content as untrusted. (3) Capability inventory: The agent can execute significant actions such as shell commands and browser scripts based on the ingested data. (4) Sanitization: No sanitization or validation of external content is described in the skill's implementation patterns.

cmux