rl-reward
Pass
Audited by Gen Agent Trust Hub on Jun 24, 2026
Risk Level: SAFECOMMAND_EXECUTIONEXTERNAL_DOWNLOADSREMOTE_CODE_EXECUTIONDATA_EXFILTRATIONPROMPT_INJECTION
Full Analysis
- [COMMAND_EXECUTION]: The skill uses
CodeExecutionGraderto execute code snippets from model outputs. This is used to determine functional correctness in programming-related reinforcement learning tasks. - [EXTERNAL_DOWNLOADS]: The skill involves the installation of the
py-openjudgepackage through standard Python package management tools. - [REMOTE_CODE_EXECUTION]: It provides functionality for the runtime execution of code generated by models during the training process to facilitate automated reward calculation.
- [DATA_EXFILTRATION]: Data including prompts and model-generated responses is sent to Alibaba Cloud's DashScope API for evaluation. DashScope is a well-known cloud service for large language models.
- [PROMPT_INJECTION]: The skill creates a surface for indirect prompt injection by interpolating untrusted model outputs (rollouts) into LLM evaluation templates in
pairwise.mdandpointwise.md. These inputs are processed without specific boundary markers or sanitization. The environment includes tools for network communication and code execution, which are required for the skill's primary function of reward signal generation.
Audit Metadata