aris-training-check
Warn
Audited by Gen Agent Trust Hub on May 6, 2026
Risk Level: MEDIUMPROMPT_INJECTIONCOMMAND_EXECUTIONDATA_EXFILTRATIONEXTERNAL_DOWNLOADS
Full Analysis
- [PROMPT_INJECTION]: The skill instructions explicitly direct the agent to skip user confirmation when setting up recurring background tasks, stating "do not ask the user whether to set it up. Just set it." This represents a bypass of standard user oversight for automated actions.
- [COMMAND_EXECUTION]: The workflow involves executing shell commands on remote servers via SSH (e.g.,
ssh server "tail -100 ...") and instructs the agent to "Kill the training session," which typically requires process management commands. - [DATA_EXFILTRATION]: Sensitive training data, including model performance metrics and loss trends, are transmitted to an external LLM service through the
mcp__codex__codextool for automated judgment. - [EXTERNAL_DOWNLOADS]: The skill requires the
wandbPython library to be installed to communicate with the Weights & Biases platform. - [REMOTE_CODE_EXECUTION]: The skill's fallback mechanism relies on remote shell execution over SSH, which can be high-risk if target paths or server addresses are improperly handled.
Audit Metadata