arkcli-train-finetune
Warn
Audited by Gen Agent Trust Hub on Oct 28, 2026
Risk Level: MEDIUMDYNAMIC_EXECUTIONEXTERNAL_DOWNLOADSCOMMAND_EXECUTIONINDIRECT_PROMPT_INJECTIONDATA_EXFILTRATION
Full Analysis
- [EXTERNAL_DOWNLOADS]: The skill instructs the agent to install the Ark SDK using pip directly from a remote URL hosted on Volcengine's infrastructure (volces.com). While this is a vendor-owned resource, remote software installation is a high-privilege operation.
- [DYNAMIC_EXECUTION]: The skill guides the agent to create, modify, and execute local Python scripts (e.g., job.py) and Reinforcement Learning plugins (rollout and grader functions). These scripts are executed using the local Python interpreter to interface with the Ark SDK.
- [COMMAND_EXECUTION]: The skill relies on numerous subprocess calls to the ArkCLI and Ark SDK binaries to manage model lifecycles, resource groups, and task monitoring.
- [INDIRECT_PROMPT_INJECTION]: The skill processes external training datasets provided by the user. Ingestion points: Local JSONL files specified via CLI arguments in references/create.md. Boundary markers: The instructions do not specify prompt-level delimiters, but the workflow utilizes the 'arkcli dataset validate' tool. Capability inventory: Local Python execution and cloud API submission via CLI. Sanitization: Relies on server-side schema validation provided by the ArkCLI dataset validation command.
- [DATA_EXFILTRATION]: As part of its intended purpose, the skill uploads local training and validation files to Volcengine's Object Storage (TOS). Users should be aware that data provided to the skill for training is transmitted to the cloud provider's environment.
Audit Metadata