The Agent Skills Directory

[PROMPT_INJECTION]: The skill instructions in SKILL.md contain directives to override environment security constraints. Specifically, the agent is instructed to 'immediately retry' a command with 'escalated permissions' if an initial attempt is blocked by sandboxing or permission failures, which is an explicit attempt to bypass security boundaries.
[COMMAND_EXECUTION]: The agent is directed to execute system-level commands (codex mcp add) to modify its own environment by installing external toolsets.
[EXTERNAL_DOWNLOADS]: The skill specifies a workflow for downloading and installing an external MCP server from a remote URL (https://developers.openai.com/mcp).
[METADATA_POISONING]: The skill provides documentation and 'official snapshots' for hypothetical and currently non-existent products (such as 'GPT-5.4', 'gpt-oss-120b', and 'Sora-2'), which presents a high risk of user deception and misinformation.

openai-docs