pytorch-model-trainer

Fail

Audited by Gen Agent Trust Hub on Feb 16, 2026

Risk Level: HIGHCOMMAND_EXECUTIONPROMPT_INJECTION
Full Analysis
  • [Indirect Prompt Injection] (HIGH): The skill defines a high-risk capability surface. It is designed to process external machine learning code, patterns, and configurations while possessing the ability to execute arbitrary python and pip commands.
  • Ingestion points: The skill ingests untrusted data whenever a user provides ML code snippets, training patterns, or data preparation requests.
  • Boundary markers: None are specified in the skill definition to separate instruction from untrusted data.
  • Capability inventory: Access to Bash(python:*), Bash(pip:*), Write, and Edit allows for full system compromise or data exfiltration if the agent is manipulated by embedded instructions in ML code.
  • Sanitization: No sanitization or validation logic is defined for the external content being processed.
  • [Command Execution] (HIGH): The skill explicitly requests permission to use Bash(python:*) and Bash(pip:*). This allows the agent to install arbitrary packages and execute arbitrary Python code, which can be leveraged for malicious persistence or lateral movement if the agent's reasoning is compromised.
Recommendations
  • AI detected serious security threats
Audit Metadata
Risk Level
HIGH
Analyzed
Feb 16, 2026, 06:57 AM