The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill provides instructions to download code from an untrusted GitHub repository (github.com/skills-directory/skill-codex.git) during installation.\n- [COMMAND_EXECUTION]: The skill enables and encourages the use of a third-party CLI (codex) with dangerous configurations, specifically --sandbox off which disables isolation and --full-auto which removes user confirmation steps. It also recommends using 2>/dev/null to suppress the tool's reasoning process from the user.\n- [REMOTE_CODE_EXECUTION]: The skill allows a remote model to perform autonomous, multi-file edits on the local system, effectively granting remote control over the codebase.\n- [INDIRECT_PROMPT_INJECTION]: The skill's primary purpose involves reading and analyzing large amounts of local code, which could contain hidden instructions that influence the behavior of the delegated model.\n- [METADATA_POISONING]: The skill lists non-existent AI models (e.g., gpt-5.5, gpt-5.4, gpt-5.3-codex-spark) and an unknown domain (ara.so), which could mislead users about the tool's capabilities and provenance.\n- [CREDENTIALS_UNSAFE]: The installation guide suggests checking and setting API keys in shell environments using commands like echo $OPENAI_API_KEY, which could lead to sensitive data exposure in shell histories.

skill-codex-delegation