Affected by GO-2026-4637
and 4 other vulnerabilities
GO-2026-4637: WeKnora has Broken Access Control - Cross-Tenant Data Exposure in github.com/Tencent/WeKnora
GO-2026-4638: WeKnora Vulnerable to Tool Execution Hijacking via Ambigous Naming Convention In MCP client and Indirect Prompt Injection in github.com/Tencent/WeKnora
GO-2026-4640: WeKnora has Unauthorized Cross‑Tenant Knowledge Base Cloning in github.com/Tencent/WeKnora
GO-2026-4641: WeKnora Vulnerable to Remote Code Execution via SQL Injection Bypass in AI Database Query Tool in github.com/Tencent/WeKnora
GO-2026-4643: WeKnora has DNS Rebinding Vulnerability in web_fetch Tool that Allows SSRF to Internal Resources in github.com/Tencent/WeKnora
FindModel looks a model name up among the vendor's entries of one model
type: exact id, then aliases, then glob patterns (longest literal prefix
wins). Matching is case-insensitive.
The type is part of the key because families share prefixes across types.
Aliyun's qwen3* chat glob also matches qwen3.8-text-embedding, and gpt-5*
matches gpt-5-embed; an untyped lookup would hand an embedding row the chat
family's compat, which the embedding overlay rejects, and the row could not
be built at all. A VLM row is a chat model that accepts images, so it looks
among the chat entries.