The Agent Skills Directory

Indirect Prompt Injection: The skill ingests untrusted data from the user (such as project IDs, email addresses, and webhook URLs) and interpolates them into Terraform HCL files and shell command arguments.
Ingestion points: User-provided values for GCP project IDs, notification channel details (email, phone, URL), and PagerDuty keys.
Boundary markers: None explicitly mentioned in the interpolation logic for the generated HCL or shell commands.
Capability inventory: The skill can write files (e.g., notification_channels.tf), execute a local bash script (scripts/verify_pubsub_topic.sh), and run gcloud and terraform CLI tools.
Sanitization: While the bundled shell script quotes arguments to prevent basic shell injection, the instructions do not specify explicit sanitization or validation of the user-provided strings before they are interpolated into the generated Terraform resources.
Command Execution: The skill executes local shell scripts and system commands to discover resources and deploy configurations.
Evidence: The main instructions and the scripts/verify_pubsub_topic.sh file utilize the gcloud CLI for project metadata retrieval, listing notification channels, and verifying IAM policies. It also invokes the terraform CLI to apply the generated infrastructure-as-code configurations.

cloud-monitoring-notification-channels-configuration