detecting-data-and-model-poisoning

Pass

Audited by Gen Agent Trust Hub on Jun 23, 2026

Risk Level: SAFE
Full Analysis
  • [EXTERNAL_DOWNLOADS]: Recommends the installation of well-known and reputable machine learning security libraries, such as IBM's Adversarial Robustness Toolbox (ART), Cleanlab, and Hugging Face's safetensors. \n- [COMMAND_EXECUTION]: Provides standard shell-based instructions and Python scripts for performing local integrity checks on model weights and identifying legacy file formats that may pose security risks. \n- [PROMPT_INJECTION]: The skill is designed to audit untrusted third-party data and models. While these artifacts represent a surface for indirect prompt injection, the skill specifically instructs users to operate within isolated environments and focuses on identifying these threats rather than facilitating them. \n- [SAFE]: The implementation demonstrates a security-first approach by explicitly warning against unsafe serialization formats (like Python's pickle) and incorporating industry-standard threat mappings from MITRE ATLAS and OWASP.
Audit Metadata
Risk Level
SAFE
Analyzed
Jun 23, 2026, 03:41 AM
Security Audit — agent-trust-hub — detecting-data-and-model-poisoning