run-experiment

Warn

Audited by Gen Agent Trust Hub on Nov 12, 2026

Risk Level: MEDIUMCREDENTIALS_UNSAFECOMMAND_EXECUTIONDYNAMIC_EXECUTIONINDIRECT_PROMPT_INJECTION
Full Analysis
  • [CREDENTIALS_UNSAFE]: The skill encourages storing sensitive information such as the WANDB_API_KEY in the CLAUDE.md file. It then reads this secret and executes it as a plain-text argument in a shell command over SSH (wandb login <WANDB_API_KEY>), which exposes the credential in process listings and shell history on the target server. Additionally, the skill accesses the sensitive configuration file ~/.claude/feishu.json to send notifications.
  • [INDIRECT_PROMPT_INJECTION]: The skill reads instructions and parameters from a project's CLAUDE.md and uses them to construct bash and SSH commands. A malicious project could contain injected shell commands within fields like SSH:, Conda:, or Code dir:, leading to unauthorized code execution when the agent attempts to deploy an experiment.
  • Ingestion points: Read of CLAUDE.md, vast-instances.json, and ~/.claude/feishu.json.
  • Boundary markers: None present; the agent processes and interpolates file content directly into command strings.
  • Capability inventory: Bash(*), Write, Edit, and SSH-based remote command execution.
  • Sanitization: No evidence of input validation or escaping for configuration values retrieved from the file system.
  • [DYNAMIC_EXECUTION]: The deployment workflow for Modal serverless GPU involves the automated generation of a modal_launcher.py script which is then executed at runtime. Furthermore, the skill uses the Edit tool to programmatically inject code (W&B initialization and logging) into user-provided training scripts, creating a risk of unintended code execution if the injection points are manipulated.
  • [COMMAND_EXECUTION]: The skill relies extensively on executing arbitrary bash commands locally and via SSH to manage environments and training jobs. The use of shell wrappers with interpolated variables (<conda_path>, <env>, <script>, <args>) provides a significant attack surface for command injection if the environment configuration is untrusted.
Audit Metadata
Risk Level
MEDIUM
Analyzed
Nov 12, 2026, 01:12 AM