This repository publishes reference data, not a guarantee that any listed service is malicious or compromised. Consumers must review policy impact before enforcement.

Please report vulnerabilities privately through GitHub's security-advisory interface when available.

The discovery workflow deliberately does not fetch newly discovered URLs. Collectors are restricted to fixed GitHub-owned API hosts so that malicious public configuration files cannot turn the workflow into an SSRF client.

The optional LLM reviewer follows the same trust boundary. It sends a bounded inventory, without raw discovered file contents, to one fixed provider endpoint. Only the selected workflow step receives that provider's API key. Model output is treated as untrusted: it is schema-checked, normalized, size-limited, marked suggestion-only, and cannot edit the catalog or generated lists. Model-proposed evidence links are stored for human inspection but are never fetched by the workflow.