huawei-cloud-cloudrobo-infer
Pass
Audited by Gen Agent Trust Hub on Sep 9, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTIONEXTERNAL_DOWNLOADSCOMMAND_EXECUTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill ingests data from external files such as skill_config.json and r2c_config.yaml downloaded from the model asset manager to drive the creation of inference services. 1. Ingestion points: Deployment parameters are fetched from asset files via the download-url API (documented in SKILL.md Step 5). 2. Boundary markers: There are no explicit boundary markers or ignore-instructions warnings for the interpolated content. 3. Capability inventory: The skill can execute service lifecycle commands (create, update, delete, start, stop) using the ingested data. 4. Sanitization: The agent is instructed to filter skill_config.json to specific fields, but r2c configuration is used as raw metadata.
- [EXTERNAL_DOWNLOADS]: The skill fetches algorithm metadata and configuration files from the vendor's official asset management domain (cloudrobo-gallery.{region}.myhuaweicloud.com) to automate parameter resolution.
- [COMMAND_EXECUTION]: The skill utilizes the cloudrobo CLI and Python SDK for infrastructure management. It includes a security control (validate_safe_id) to protect against path traversal attacks when handling service identifiers.
Audit Metadata