Security model

AI Model Directory is provider-supplied metadata. It is useful for model selection, defaults, and comparison, but it should not be treated as a billing authority or a guarantee that a provider is telling the truth.

What the updater protects

• Provider endpoints are hardcoded in source.
• Fetched data is validated against a strict Zod schema before writing files.
• Model IDs are normalized into safe directory names.
• Conflicting normalized model directories are resolved deterministically.
• Terminal output is sanitized before logging.
• Network fetches have a 60 second timeout.
• IDs and names are length-limited and reject raw control characters.
• Removed upstream models are deleted unless metadata.toml sets preserve = true.
• extends paths are restricted to paths inside data/providers.
• Provider-supplied code, shell commands, and HTML are not executed.

What your app should still do

• Treat prices and capabilities as a strong default, not a source of truth.
• Keep provider API keys in your own backend or secret store.
• Handle missing models and changed limits gracefully.
• Check provider billing pages for critical cost decisions.

Reporting issues

If a provider entry looks suspicious or incorrect, open an issue or pull request with a link to the provider's official docs.