The Agent Skills Directory

[COMMAND_EXECUTION]: The skill explicitly instructs users to launch sessions using the --dangerously-skip-permissions flag with the cmux tool. This bypasses the security permission prompts required for agent actions, allowing the agent to perform potentially destructive or unauthorized operations without user confirmation.
[COMMAND_EXECUTION]: The documentation describes the use of lifecycle hooks (TaskCompleted, TaskCreated, TeammateIdle) configured in settings.json that execute arbitrary shell commands. This creates a persistent dynamic execution mechanism where system commands are triggered automatically by agent state changes.
[COMMAND_EXECUTION]: The skill provides examples of spawning multiple agents that have the capability to execute complex build and testing tools (e.g., xcodebuild, pnpm test, pnpm typecheck) across various local repositories and system directories.
[PROMPT_INJECTION]: The skill includes specific instructions on how to phrase prompts to trigger the experimental agent-teams feature (e.g., "Say 'spawn team agents' or 'create an agent team', not 'subagents'"). While intended for feature activation, these instructions demonstrate how to influence agent behavior through specific verbal triggers.
[DATA_EXFILTRATION]: While no explicit exfiltration was detected, the skill facilitates 'Live incident response' and 'Overnight research' scenarios where agents are instructed to ingest production logs, bug reports, and external documentation. This establishes an attack surface for indirect prompt injection if the ingested data contains malicious instructions, which could then be executed via the aforementioned command execution capabilities.

agent-teams