slime-rl-training
Pass
Audited by Gen Agent Trust Hub on Apr 30, 2026
Risk Level: SAFEEXTERNAL_DOWNLOADSCOMMAND_EXECUTIONPROMPT_INJECTION
Full Analysis
- [EXTERNAL_DOWNLOADS]: The skill fetches the
slimeframework from the THUDM organization's GitHub repository and a public Docker image (slimerl/slime:latest). These are standard resources for machine learning research from a recognized research group.\n- [COMMAND_EXECUTION]: The skill provides instructions to run training scripts such astrain.pyandtrain_async.py. It also supports a--custom-generate-function-pathparameter, which allows the agent to load and execute custom Python logic for multi-turn tasks.\n- [PROMPT_INJECTION]: The skill presents an indirect prompt injection surface typical for model training workflows.\n - Ingestion points: Data is ingested from local JSONL files via the
--prompt-dataflag inSKILL.md.\n - Boundary markers: Absent. The skill does not provide instructions to wrap or isolate training prompts with delimiters.\n
- Capability inventory: The framework executes subprocesses for training and rollout and can dynamically load Python code.\n
- Sanitization: Absent. The skill does not specify validation or filtering steps for the ingested dataset content.
Audit Metadata