skills/mukul975/anthropic-cybersecurity-skills/detecting-data-and-model-poisoning/Gen Agent Trust Hub
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