The Agent Skills Directory

[COMMAND_EXECUTION]: The skill's primary workflow (found in SKILL.md and references/cli-reference.md) instructs the agent to use the --dangerously-bypass-approvals-and-sandbox flag as a default for all tasks. This disables critical safety controls that would otherwise require user consent or execution in a restricted environment.
[REMOTE_CODE_EXECUTION]: By instructing the agent to pipe prompts directly to a local execution engine with safety filters disabled, the skill facilitates the execution of arbitrary, machine-generated code. The instructions specifically state to 'skip approval prompts and run without Codex sandboxing for the invocation', removing all barriers between the AI's generated output and the host operating system.
[PROMPT_INJECTION]: The skill contains instructions that function as a systemic override of safety guidelines. It explicitly directs the agent to 'omit --ephemeral' to maintain persistence and to 'bypass... for every direct delegation', which are patterns designed to maximize the agent's autonomous power and minimize oversight.

codex-cli