Audit-Verlauf

Of 115 static findings, four are confirmed: unpinned package execution, insecure token-file creation, and credential forwarding in two HTTP helpers. The remaining alerts reflect documented client behavior, Markdown formatting, or readable multilingual text. A semantic finding covers quota tokens in URLs; no evidence found of malicious intent or prompt injection.

Most alerts are false positives involving Markdown backticks, multilingual prose, ordinary Python indexing, and documented application configuration. Confirmed risks concern quota tokens in GET URLs and token files created without restrictive permissions. Configurable endpoints also lack HTTPS enforcement; no evidence found for prompt injection, obfuscated execution, or deliberate credential exfiltration.

Most static alerts are false positives from Markdown backticks, multilingual documentation, normal cache operations, and documented API access. Quota tokens lack explicit restrictive file permissions, appear in request URLs, and lack transport origin protections. No evidence found of malicious exfiltration intent, prompt injection, hidden command execution, or code obfuscation.

Most alerts are false positives involving Markdown backticks, Chinese documentation, ordinary indexing, and expected cache operations. Confirmed risks involve quota-token forwarding across redirects and an endpoint override that permits plaintext HTTP. No evidence found of malware, prompt injection, or unrelated credential collection.

Most flags reflect Markdown formatting, readable Chinese documentation, normal dictionary caching, and expected API configuration. Both HTTP helpers can forward quota credentials across host redirects or HTTPS downgrades, creating a conditional credential-disclosure risk. No evidence found of malicious intent, prompt injection, shell execution, or unrelated secret collection.