simpo-training

Pass

Audited by Gen Agent Trust Hub on Sep 9, 2026

Risk Level: SAFEEXTERNAL_DOWNLOADSINDIRECT_PROMPT_INJECTION
Full Analysis
  • [EXTERNAL_DOWNLOADS]: The skill provides instructions to download the official alignment-handbook repository from HuggingFace's GitHub account and to install common machine learning dependencies from standard package registries. These sources are well-known and considered trusted.
  • [INDIRECT_PROMPT_INJECTION]: The skill defines workflows for processing preference datasets from external hubs. These datasets represent a surface for indirect prompt injection as they contain untrusted user-generated content.
  • Ingestion points: External datasets defined in the dataset_mixer field within the YAML training configurations (e.g., UltraFeedback, HelpSteer).
  • Boundary markers: Not explicitly defined in the provided configuration files.
  • Capability inventory: The skill facilitates model training, which involves file writes to an output directory and execution of training scripts via the accelerate CLI.
  • Sanitization: While documentation provides examples for data quality filtering (length and diversity), there is no explicit sanitization for adversarial prompt injection strings.
Audit Metadata
Risk Level
SAFE
Analyzed
Sep 9, 2026, 07:07 PM
Security Audit — agent-trust-hub — simpo-training