The Agent Skills Directory

[COMMAND_EXECUTION]: The skill documents the use of cmux send and cmux new-workspace --command, which allow the agent to send arbitrary strings to terminal surfaces. This capability enables the execution of shell commands on the local system through the cmux CLI.
[DATA_EXFILTRATION]: The agent is provided with tools to access and retrieve potentially sensitive data from the user's workspace environment. The cmux read-screen command allows the agent to capture text content from terminal panes, while cmux browser snapshot and cmux browser screenshot allow it to extract text and visual data from active browser surfaces.
[INDIRECT_PROMPT_INJECTION]: The skill facilitates the processing of untrusted data from terminal outputs and web pages, creating a surface for indirect prompt injection attacks.
Ingestion points: Untrusted data enters the agent context via cmux read-screen (terminal history) and cmux browser snapshot (web page content).
Boundary markers: The instructions do not define boundary markers or delimiters to differentiate between application data and agent instructions.
Capability inventory: The agent possesses high-impact capabilities, such as cmux send for command execution and cmux rpc for direct application state manipulation, which could be triggered by malicious content found on a screen or web page.
Sanitization: There is no mention of sanitizing, filtering, or validating the data retrieved from the environment before it is processed or acted upon by the agent.

cmux