slime-rl-training
Pass
Audited by Gen Agent Trust Hub on Oct 29, 2026
Risk Level: SAFEEXTERNAL_DOWNLOADSCOMMAND_EXECUTIONDYNAMIC_EXECUTIONINDIRECT_PROMPT_INJECTION
Full Analysis
- [EXTERNAL_DOWNLOADS]: In SKILL.md, the instructions guide the user to clone the 'slime' framework from the official THUDM GitHub repository (https://github.com/THUDM/slime.git) and pull the project's Docker image (slimerl/slime:latest). These are legitimate and well-known sources for the framework described.
- [COMMAND_EXECUTION]: SKILL.md and troubleshooting.md include various shell commands for environment setup, sourcing model configurations from scripts/models/, and launching training processes with train.py and train_async.py. These are standard operations for machine learning research.
- [DYNAMIC_EXECUTION]: As detailed in SKILL.md and api-reference.md, the framework supports loading and executing user-supplied Python modules for custom reward calculations and multi-turn generation logic (e.g., via the --custom-generate-function-path flag). This is a core intended feature for customizing reinforcement learning behavior.
- [INDIRECT_PROMPT_INJECTION]: The skill documents an attack surface for indirect prompt injection common to training frameworks. 1. Ingestion points: The --prompt-data argument in SKILL.md specifies where external datasets enter the agent context. 2. Boundary markers: No explicit delimiters or 'ignore' instructions for data ingestion are present in the examples. 3. Capability inventory: The framework supports executing tool calls based on generated content as shown in Workflow 3. 4. Sanitization: No explicit sanitization or validation of input data is documented in the provided code snippets.
Audit Metadata