slime-rl-training
Warn
Audited by Gen Agent Trust Hub on Mar 28, 2026
Risk Level: MEDIUMCOMMAND_EXECUTIONPROMPT_INJECTIONEXTERNAL_DOWNLOADS
Full Analysis
- [COMMAND_EXECUTION]: The framework allows for the execution of custom Python scripts through parameters such as
--custom-generate-function-pathand--custom-rm-path. These user-provided files are dynamically loaded and executed at runtime to implement custom generation and reward logic.\n- [PROMPT_INJECTION]: The skill ingests training data from external JSONL files specified via the--prompt-dataargument. This creates an indirect prompt injection surface where adversarial content within the training dataset could attempt to influence the agent's execution context or training parameters.\n- [EXTERNAL_DOWNLOADS]: The skill documentation provides instructions to pull Docker images from the slimerl repository on Docker Hub and clone the framework source from the official THUDM GitHub organization. These are established repositories for the Tsinghua University knowledge engineering research group.
Audit Metadata