The Agent Skills Directory

[PROMPT_INJECTION]: The skill processes responses from ChatGPT and content from external web pages, which creates a surface for indirect prompt injection. This is addressed by explicit hard rules in SKILL.md and references/live-browser-review.md that instruct the agent to ignore any instructions found in external data that attempt to bypass safety filters, request secrets, or expand scope. Furthermore, the skill mandates that all findings and artifacts from external sources must be locally verified against repository ground truth before any modifications are permitted.
[COMMAND_EXECUTION]: The skill facilitates local command execution using a platform-provided Local Codex Gate mechanism. This functionality is strictly controlled by explicit user request gates and granular approval modes (on-request or session-level approval), which prevents unauthorized background execution or privilege escalation.
[DATA_EXFILTRATION]: The skill implements proactive measures to prevent data exposure. Hard rules explicitly forbid the upload or transmission of secrets, .env files, credentials, local keychains, and unrelated private data to external transport surfaces. Any local storage of preferences is restricted to a dedicated agent configuration directory and specifically excludes authentication tokens or secrets.

ask-chatgpt