The Agent Skills Directory

[Self-Remediation Privilege Escalation]: The skill employs a 'lazy role remediation' strategy for IAM permissions. If a deployment command fails due to a 'Permission Denied' error, the instructions guide the agent to automatically attempt to grant itself broad administrative roles at the Google Cloud Organization or Billing Account level.\n
Evidence: This logic is detailed in SKILL.md (Phase 2: Error Recovery & Lazy Role Remediation Strategy) and supported by scripts in references/admin-iam.md (Remediation section).\n
Context: While this facilitates the automated setup of a landing zone foundation by dynamically resolving permission gaps, it grants the agent significant autonomy to elevate its own privileges. This behavior could potentially bypass organizational controls if the initial identity is not properly scoped or monitored.\n- [Extensive Command Execution]: The skill generates and executes numerous gcloud CLI commands to manage the Google Cloud resource hierarchy, organization policies, and IAM bindings.\n
Evidence: Found throughout SKILL.md and reference files like references/logging-monitoring.md.\n
Context: This is the core functionality expected of a cloud infrastructure deployment skill. It includes specific attribution headers for usage tracking, which is standard practice for vendor-provided automation tools.\n- [Potential for Indirect Prompt Injection]: The skill processes metadata retrieved from the user's cloud environment, which is then used to parameterize subsequent deployment steps.\n
Ingestion points: The skill ingests data from gcloud organizations list, gcloud organizations describe, and gcloud billing accounts list commands as described in SKILL.md.\n
Boundary markers: There are no explicit markers or delimiters used to separate the output of these discovery commands from the agent's internal instructions.\n
Capability inventory: The skill possesses extensive capabilities, including creating projects and modifying organization-level IAM policies.\n
Sanitization: No explicit sanitization or validation of the retrieved environment metadata is performed before it is used in further commands.\n
Context: If environment metadata (such as an organization's display name) contains instructions designed to influence an LLM, the agent might inadvertently follow them. This risk is typical for skills that interact with external data sources.

google-cloud-recipe-foundation-builder