gptq

Pass

Audited by Gen Agent Trust Hub on Feb 17, 2026

Risk Level: SAFEEXTERNAL_DOWNLOADSCOMMAND_EXECUTIONPROMPT_INJECTION
Full Analysis
  • Unverifiable Dependencies (LOW): The guides suggest loading datasets from third-party Hugging Face accounts (e.g., 'anon8231489123/ShareGPT_Vicuna_unfiltered'). Loading data from unverified sources is a potential security concern as the content is not controlled by the model or tool developer.
  • Privilege Escalation (LOW): Troubleshooting documentation includes instructions for using 'sudo' to install system dependencies. While typical for environment setup, this identifies a request for elevated permissions within the user's instructions.
  • Indirect Prompt Injection (LOW): The skill identifies a potential surface for indirect prompt injection by demonstrating how to ingest external datasets for model calibration without implementing sanitization or boundary markers.
  • Ingestion points: 'load_dataset' function calls throughout 'references/calibration.md'.
  • Boundary markers: None identified in the provided implementation examples.
  • Capability inventory: The skill documents capabilities including shell command execution via Docker ('integration.md') and file system writes ('calibration.md').
  • Sanitization: No sanitization or validation of dataset content is demonstrated before it is used to calibrate model weights.
Audit Metadata
Risk Level
SAFE
Analyzed
Feb 17, 2026, 04:58 PM