The Agent Skills Directory

[METADATA_POISONING]: The skill documentation and model list contain deceptive claims regarding supported OpenAI models.
Evidence: The skill claims to support 'gpt-5.2', 'gpt-5.3-codex', 'gpt-5-mini', and 'gpt-4.1', none of which are officially released OpenAI models.
Evidence: It incorrectly labels established models like 'gpt-4o' and 'gpt-4o-mini' as 'legacy'.
[COMMAND_EXECUTION]: The skill uses shell commands and subprocesses for setup and runtime functionality.
Evidence: The setup instructions require executing cp -r, pnpm install, and pnpm run commands.
Evidence: The server-side audio utility in templates/lib/integrations-openai-ai-server/src/audio/client.ts spawns ffmpeg as a subprocess to convert audio formats.
[EXTERNAL_DOWNLOADS]: The skill relies on external package registries and automated setup tools.
Evidence: Setup instructions involve calling setupReplitAIIntegrations and installing several Node.js dependencies (openai, p-limit, p-retry) from public registries.
[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted user input which is later passed to LLM providers.
Ingestion points: User-provided audio data and text messages processed by the /openai/conversations endpoints.
Boundary markers: The template does not implement explicit delimiters or system instructions to ignore potentially malicious content embedded in user inputs.
Capability inventory: The skill has the ability to execute shell commands (via ffmpeg), read/write to the filesystem (temporary audio files), and perform network operations through the OpenAI SDK.
Sanitization: Audio input is converted to WAV format, but no natural language sanitization is evident in the provided message handling logic.

ai-integrations-openai