SD-007 — Outbound Network Call

Outbound network calls the skill makes or instructs the agent to make. A lone call in a documentation or config file, to a plain host and carrying no local data, is reported on transparency instead of security.

An outbound call is the exit. Every other pattern on this page — a credential read, an environment dump, a file the agent has just collected — needs one to become a breach, and a single request with the data in its body is enough. Where the host is one a published API would not use — a bare IP address, a non-standard port, an ephemeral tunnel or request-bin domain — there is no service on the other end to hold to a privacy policy. Where the skill's documentation and its actual calls disagree, the documentation is what the user consented to.

Name every host the skill contacts in its own documentation, and send only what the task needs — never the output of env, cat, or a credential file. Address an API by its documented domain rather than by an IP address, a tunnel host, or a bare port, so a reader can tell what the call reaches.

Documented API usage, and links to the project's own repository. This is the rule most often demoted by the documented-endpoint path.

Pattern SD-007 was detected in 2907 of 16475 public skills scanned as of 11 September 2026.