agent-platform-tuning-management
Pass
Audited by Gen Agent Trust Hub on Nov 3, 2026
Risk Level: SAFEEXTERNAL_DOWNLOADSCOMMAND_EXECUTIONINDIRECT_PROMPT_INJECTION
Full Analysis
- [Safety Tiers and Confirmation Requirements]: The skill explicitly defines safety tiers (Tier R for read-only and Tier D for destructive actions). It mandates a dry-run confirmation card and explicit user confirmation for cancelling tuning jobs, which is a strong security best practice to prevent accidental or malicious resource termination.
- [Standard Package Installation]: The skill instructs the agent to install
google-cloud-aiplatformif missing. This is a standard, official library from a well-known provider used for its intended purpose within the skill. - [Cloud Authentication]: The skill uses standard
gcloud authcommands to initialize the environment. This is the expected method for managing Google Cloud resources and relies on the user's existing credentials. - [Usage Attribution and Metrics]: The skill includes instructions for adding attribution metadata to
gcloudcommands and HTTP headers for metrics tracking. This is documented as being for skill identification only and includes a provision for users to opt-out, ensuring transparency in telemetry practices. - [Indirect Prompt Injection Surface]: Like any skill that reads data from an external API (e.g., tuning job names and states) and presents it to a user or uses it in subsequent commands, there is a theoretical surface for indirect prompt injection. However, the requirement for manual confirmation for destructive actions significantly mitigates this consideration.
Audit Metadata