slime-rl-training
Pass
Audited by Gen Agent Trust Hub on Sep 15, 2026
Risk Level: SAFEEXTERNAL_DOWNLOADSDYNAMIC_EXECUTIONCOMMAND_EXECUTIONINDIRECT_PROMPT_INJECTION
Full Analysis
- [EXTERNAL_DOWNLOADS]: The skill instructs users to pull a Docker image from a public registry and clone the slime framework source code from a GitHub repository.
- [DYNAMIC_EXECUTION]: The framework allows for the execution of custom logic by loading user-provided Python scripts through the
--custom-generate-function-pathand--custom-rm-patharguments. - [COMMAND_EXECUTION]: The skill utilizes shell commands to initiate training runs, source model-specific configurations, and manage the execution environment.
- [INDIRECT_PROMPT_INJECTION]: The skill ingests training data from external JSONL files, which represents a potential attack surface if the data contains malicious instructions designed to influence the training process.
- Ingestion points: Training data is read from paths specified by the
--prompt-dataargument in the training launch commands. - Boundary markers: No explicit delimiters or instructions to ignore embedded prompts are mentioned in the skill files.
- Capability inventory: The skill has the capability to execute shell commands for training and load external Python code for reward modeling.
- Sanitization: There is no evidence of input validation or sanitization for the training datasets described in the documentation.
Audit Metadata