The Agent Skills Directory

[COMMAND_EXECUTION]: The skill relies on the codex exec tool to spawn subagents that can run arbitrary shell commands, perform deep codebase exploration, and execute multi-step workflows.
[PRIVILEGE_ESCALATION]: Multiple examples in SKILL.md utilize the --dangerously-bypass-approvals-and-sandbox flag, which is a high-risk parameter designed to circumvent built-in security approvals and sandbox protections for the subagent's execution environment.
[PROMPT_INJECTION]: The skill provides instructions for the subagent to 'Act autonomously, no permission asking' and 'Make decisions and proceed boldly', which are patterns used to override default agent safety and autonomy constraints.
[INDIRECT_PROMPT_INJECTION]: The skill creates a surface for indirect prompt injection by spawning agents with significant capabilities (e.g., web search, file access, sandbox bypass) that ingest user-defined task prompts without evidence of input sanitization or explicit boundary markers.
[DYNAMIC_EXECUTION]: The scripts/codex-parent-settings.sh script dynamically generates and executes a Python script via a heredoc to parse the ~/.codex/config.toml file, while the skill also executes a PowerShell script (codex-parent-settings.ps1) to perform similar configuration extraction.

codex-subagent