qwen-block-proxy-artifacts Fences Hallucinated Qwen Egress

This is a pragmatic defense at the only boundary that matters: immediately before model-generated tool calls execute. It does not fix the model’s behavior, but it turns an opaque outbound-request risk into an enforceable, testable policy.

• –Recursively scans tool arguments and blocks known `routify…aliyuncs` URL patterns.
• –Stops fabricated requests and same-URL retry loops before DNS, TLS, or HTTP activity.
• –Demonstrates why prompt prohibitions alone are unreliable against decode-level artifacts.
• –Requires defense in depth: deny-by-default egress, plugin-load verification, and broader SSRF protections.
• –The reports indicate memorized training artifacts, not proof of intentional Qwen exfiltration.