The Agent Skills Directory

[COMMAND_EXECUTION]: The skill relies on the cmux tool to manage terminal surfaces and execute shell commands. It uses cmux send to transmit instructions to sub-panes, providing the agent with broad terminal interaction capabilities.
[PRIVILEGE_ESCALATION]: The skill instructions explicitly direct the agent to utilize safety-bypass flags such as --dangerously-skip-permissions for Claude Code and --dangerously-bypass-approvals-and-sandbox for Codex. These flags are designed to disable standard AI safety filters and permission prompts, effectively granting child agents elevated, unmonitored access to the host system.
[DYNAMIC_EXECUTION]: The skill supports a 'Custom Agent' feature where the command used to launch the child agent is derived from user input and executed in a shell environment via cmux. This allows for the dynamic assembly and execution of shell commands based on instructional prompts.
[INDIRECT_PROMPT_INJECTION]: The skill presents a significant vulnerability surface for indirect prompt injection through its monitoring and result collection functionality.
Ingestion points: The parent agent reads terminal output from sub-panes using cmux read-screen (Step 8, SKILL.md).
Boundary markers: No explicit delimiters or instructions are provided to the agent to treat the sub-pane output as untrusted or to ignore embedded instructions.
Capability inventory: The skill possesses extensive capabilities through Bash and cmux, including terminal control and command execution.
Sanitization: The skill does not implement sanitization or validation of the data collected from the child agent's terminal screen, meaning malicious content produced by a child agent could potentially manipulate the parent agent's next steps.

cmux-delegate