run-experiment

Warn

Audited by Gen Agent Trust Hub on Apr 5, 2026

Risk Level: MEDIUMCOMMAND_EXECUTIONREMOTE_CODE_EXECUTIONEXTERNAL_DOWNLOADSDATA_EXFILTRATIONPROMPT_INJECTION
Full Analysis
  • [COMMAND_EXECUTION]: The skill uses the Bash tool to execute various commands for environment detection, GPU monitoring (nvidia-smi), and job management (screen) both locally and on remote targets.\n- [REMOTE_CODE_EXECUTION]: Deploys and executes training scripts on remote servers via SSH and on rented Vast.ai instances.\n- [EXTERNAL_DOWNLOADS]: Automatically triggers the installation of Python dependencies from a project's requirements.txt file on remote instances using pip.\n- [DATA_EXFILTRATION]: Accesses potentially sensitive configuration files, including project-level CLAUDE.md and ~/.claude/feishu.json, which may contain sensitive API keys or connection details as described in the skill instructions.\n- [PROMPT_INJECTION]: The skill is vulnerable to indirect prompt injection due to its processing of untrusted project data. Specifically: \n
  • Ingestion points: Reads project-specific files like CLAUDE.md, vast-instances.json, requirements.txt, and training scripts (SKILL.md).\n
  • Boundary markers: None implemented for the data read from project files.\n
  • Capability inventory: Uses Bash(*), Write, Edit, and SSH access to execute arbitrary commands and modify files (SKILL.md).\n
  • Sanitization: No escaping or validation is performed on the content extracted from the ingested files before using it in shell commands or script modifications.
Audit Metadata
Risk Level
MEDIUM
Analyzed
Apr 5, 2026, 05:07 AM