arkcli-train-finetune

Pass

Audited by Gen Agent Trust Hub on Aug 28, 2026

Risk Level: SAFEEXTERNAL_DOWNLOADSREMOTE_CODE_EXECUTIONDYNAMIC_EXECUTIONINDIRECT_PROMPT_INJECTION
Full Analysis
  • [EXTERNAL_DOWNLOADS]: The skill fetches the official Ark SDK from a vendor-controlled distribution endpoint (https://ark-public-example-cn-beijing.tos-cn-beijing.volces.com/ark-sdk/ark_sdk-latest.tar.gz).
  • [REMOTE_CODE_EXECUTION]: It installs the downloaded package via pip and executes its commands (ark) to interact with the model customization service.
  • [DYNAMIC_EXECUTION]: The skill initializes project workspaces, generates Python configuration scripts (job.py) based on templates, and executes them to submit training jobs.
  • [INDIRECT_PROMPT_INJECTION]: The skill processes external data including training files, TOS URLs, and dataset references. Ingestion points are defined in SKILL.md and references/create.md. While it lacks explicit boundary markers for all data types, it performs schema validation and is limited to the capability tier of a management tool.
Audit Metadata
Risk Level
SAFE
Analyzed
Aug 28, 2026, 02:14 AM
Security Audit — agent-trust-hub — arkcli-train-finetune