Audit-Verlauf

Most static alerts match Markdown, readable Chinese prose, or a credential check that reports only presence. The CLI examples execute an unpinned npm package with access to the agent environment. Polling lacks a time or attempt limit; no evidence found of credential exfiltration, encoded payloads, or audit-directed prompt injection.

Of 54 static alerts, 53 are false positives involving Markdown, ordinary Chinese text, service links, or documented authentication. One alert covers real execution of an unpinned npm package; a separate semantic finding identifies polling without a deadline or attempt limit. No evidence found of prompt injection, credential exfiltration, or reconnaissance in the supplied files; external package and service behavior was not verified.