The Agent Skills Directory

[COMMAND_EXECUTION]: The skill uses the cmux send command to execute arbitrary shell commands in separate terminal panes, providing the agent with broad execution capabilities within the local environment.
[REMOTE_CODE_EXECUTION]: Instructions in SKILL.md direct the agent to launch sub-agents (Claude Code and Codex) using flags that bypass security approvals and sandboxes, specifically --dangerously-skip-permissions and --dangerously-bypass-approvals-and-sandbox. This effectively creates unconstrained execution environments that the main agent can interact with.
[PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection due to its core functionality of reading untrusted external data.
Ingestion points: The agent is instructed to read terminal output from other panes (cmux read-screen in SKILL.md) and extract content from web pages (cmux browser snapshot or get text in references/browser.md).
Boundary markers: There are no instructions to use delimiters or ignore directives when processing data from these external sources.
Capability inventory: The skill has high-privilege capabilities, including shell access via cmux send and the ability to manage sub-agents.
Sanitization: The skill does not implement any validation or filtering of the content ingested from terminals or the browser.
[DATA_EXFILTRATION]: The integrated browser automation tools, such as cmux browser snapshot, cookies get, and screenshot, enable the agent to access and potentially capture sensitive user data from web sessions.

cmux