slime-rl-training
Pass
Audited by Gen Agent Trust Hub on Sep 5, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTIONDYNAMIC_EXECUTIONEXTERNAL_DOWNLOADS
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill ingests training data from external JSONL files via the --prompt-data parameter, creating a potential surface for indirect prompt injection. Ingestion points: Data is loaded from files specified by the --prompt-data flag in training commands. Boundary markers: The provided templates lack explicit delimiters or instructions to ignore embedded prompts in the source data. Capability inventory: The skill executes subprocesses to launch model training and performs filesystem operations for logging and checkpoint storage. Sanitization: The documentation does not describe any validation or filtering mechanisms for prompt content.
- [DYNAMIC_EXECUTION]: The framework supports loading and running arbitrary Python logic for custom reward modeling and multi-turn generation using the --custom-generate-function-path and --custom-rm-path arguments.
- [EXTERNAL_DOWNLOADS]: The skill documentation includes instructions to download code and container images from the THUDM GitHub repository and the slimerl Docker Hub repository, which are resources associated with an established research project.
Audit Metadata