qwen-agentworld-language-world-model

MEDIUM W011: Third-party content exposure detected (indirect prompt injection risk).

• Third-party content exposure detected (high risk: 0.75). The required runtime workflow ingests benchmark JSONL fields like system_str and current_prompt (from AgentWorldBench/*_test.jsonl) into the model’s chat messages; those dataset texts are outsider-authored free-form prompts/trajectories, creating indirect prompt-injection risk.

MEDIUM W012: Unverifiable external dependency detected (runtime URL that controls agent).

• Potentially malicious external URL detected (high risk: 0.90). The vLLM deployment example explicitly runs vllm serve with --trust-remote-code while loading the HuggingFace model repo Qwen/Qwen-AgentWorld-35B-A3B (https://huggingface.co/Qwen/Qwen-AgentWorld-35B-A3B), which causes remote repository code to be fetched and executed at runtime.