The Agent Skills Directory

[COMMAND_EXECUTION]: The skill facilitates the execution of high-privilege system operations.
Evidence: The documentation specifies that the --sandbox danger-full-access flag is used for "Installing packages, running tests, system operations."
[COMMAND_EXECUTION]: The skill allows for autonomous actions without user oversight.
Evidence: The --full-auto flag is explicitly designed to "Allow file edits without confirmation prompts" and for "unattended execution" when combined with sandbox levels.
[REMOTE_CODE_EXECUTION]: The skill provides a mechanism for arbitrary task execution based on natural language.
Evidence: The codex exec "task description" command allows the agent to execute complex tasks which, when paired with high-privilege sandbox levels, can result in arbitrary code execution on the host system.
[PROMPT_INJECTION]: The skill processes external, potentially untrusted data that could contain malicious instructions.
Ingestion points: The skill ingests uncommitted changes, branch diffs, and specific commit data via codex review.
Boundary markers: No boundary markers or "ignore previous instructions" warnings are included in the skill definition to protect against instructions embedded in the code being reviewed.
Capability inventory: The agent has the capability to write to the workspace (workspace-write) and execute system commands (danger-full-access).
Sanitization: There is no evidence of input sanitization or validation performed on the code content before it is processed by the AI.

codex