The Agent Skills Directory

[COMMAND_EXECUTION]: The skill provides bash command templates for running the codex exec tool. These commands frequently use 2>/dev/null to suppress error output, which can mask critical security warnings or diagnostic failures from the host system during execution.
[REMOTE_CODE_EXECUTION]: The skill enables an external agent to perform tasks within the local workspace. When using the --sandbox workspace-write flag, the remote agent is granted permissions to write and modify files in the repository. The documentation also references a danger-full-access mode which removes sandbox protections entirely.
[DATA_EXFILTRATION]: Usage of the CLI tool involves transmitting codebase context, including directory structures and file contents, to the external AI model provider. The skill also documents a --search flag that allows the agent to perform live web searches, potentially enabling data transmission to arbitrary external domains.
[PROMPT_INJECTION]: The skill is vulnerable to indirect prompt injection as it ingests data from local repositories and feeds it into an external reasoning model.
Ingestion points: Files located within the directory path specified by the -C flag (described in SKILL.md and references/invocation.md) are processed as context.
Boundary markers: The skill does not implement or recommend specific delimiters or boundary markers to distinguish untrusted repository content from the agent's instructions.
Capability inventory: The toolchain includes the ability to read/write files and perform network operations via the codex CLI.
Sanitization: There is no programmatic sanitization of ingested content; however, the skill advises users to 'verify claims against the code' and 'review any diffs' before implementation.
[PROMPT_INJECTION]: The skill contains role-play instructions ('If you are Codex, you do not need this skill') and references to speculative or non-existent model versions such as 'gpt-5.5' and future dates, which may lead to unpredictable agent behavior when processing these instructions.

use-codex