qwen-agentworld-language-world-model
Pass
Audited by Gen Agent Trust Hub on Jun 27, 2026
Risk Level: SAFEEXTERNAL_DOWNLOADSREMOTE_CODE_EXECUTIONCOMMAND_EXECUTIONPROMPT_INJECTION
Full Analysis
- [EXTERNAL_DOWNLOADS]: The skill downloads large-scale model weights and evaluation benchmarks from Hugging Face (huggingface.co/Qwen) and ModelScope (modelscope.cn).
- [REMOTE_CODE_EXECUTION]: The deployment instructions for vLLM include the '--trust-remote-code' flag. This allows the execution of arbitrary Python code defined within the model's configuration file on the local system.
- [COMMAND_EXECUTION]: The installation and deployment sections require the user to execute multiple shell commands, including environment setup, dependency installation, and server launching.
- [PROMPT_INJECTION]: The skill exhibits an indirect prompt injection surface because it ingests untrusted data from external simulation files and agent trajectories to generate environment states.
- Ingestion points: Data enters the system via JSONL files in the 'AgentWorldBench' directory and through agent-generated actions during simulation loops.
- Boundary markers: No explicit delimiters or warnings to ignore embedded instructions are present in the provided system prompt templates.
- Capability inventory: The skill uses subprocess execution (via vllm/sglang), network operations (via the OpenAI client), and local file writing for trajectory logging.
- Sanitization: No evidence of input validation or sanitization is present for the data processed from evaluation files or agent actions.
Audit Metadata