qwen-agentworld-language-world-model
Pass
Audited by Gen Agent Trust Hub on Oct 8, 2026
Risk Level: SAFEEXTERNAL_DOWNLOADSCOMMAND_EXECUTIONREMOTE_CODE_EXECUTIONINDIRECT_PROMPT_INJECTION
Full Analysis
- [EXTERNAL_DOWNLOADS]: The skill downloads model weights (Qwen/Qwen-AgentWorld-35B-A3B) and benchmark datasets (Qwen/AgentWorldBench) from the Hugging Face Hub using the huggingface-cli tool.
- [COMMAND_EXECUTION]: The skill provides commands for installing core dependencies (torch, transformers, sglang, vllm) and launching model servers via Python subprocesses.
- [REMOTE_CODE_EXECUTION]: The vLLM deployment command includes the --trust-remote-code flag, which allows the model repository to execute custom code for its architecture. This is a common practice for newer model architectures from established vendors like Qwen/Alibaba.
- [INDIRECT_PROMPT_INJECTION]: The skill creates a vulnerability surface by processing untrusted task descriptions and trajectories from benchmark files (JSONL) and simulating environment responses based on user-supplied actions.
- Ingestion points: AgentWorldBench data files (e.g., terminal_test.jsonl) and user interaction loops in simulation scripts.
- Boundary markers: None present in the provided implementation examples.
- Capability inventory: Network access via OpenAI API client and command execution for model serving.
- Sanitization: No sanitization or validation of the input strings is shown in the deployment examples.
Audit Metadata