The Agent Skills Directory

[COMMAND_EXECUTION]: The initialization script scripts/tf-init-gcp.sh interpolates command-line arguments directly into shell commands such as mkdir and cd, as well as within unquoted heredocs. This structure allows for command injection if the input parameters are maliciously crafted.
[INDIRECT_PROMPT_INJECTION]: The skill possesses an indirect injection surface through the scripts/tf-init-gcp.sh script which processes untrusted input.
Ingestion points: Command-line arguments ($1, $2, $3) in scripts/tf-init-gcp.sh.
Boundary markers: None present to delimit user-supplied input from the shell execution environment.
Capability inventory: File system modification (mkdir), directory navigation (cd), file creation/writing (cat), and external process invocation (terraform init).
Sanitization: The script only checks for empty strings but does not sanitize, escape, or validate the content of the input arguments before interpolation.
[DYNAMIC_EXECUTION]: The skill uses scripts/tf-init-gcp.sh to dynamically generate infrastructure-as-code files (main.tf, variables.tf, outputs.tf, terraform.tfvars) and immediately proceeds to execute terraform init. This pattern of generating executable code from variables and then running it introduces a risk of local code execution.
[COMMAND_EXECUTION]: The Terraform template provided in assets/vpc-module.tf includes a firewall rule definition google_compute_firewall.allow_ssh that permits inbound TCP traffic on port 22 from the entire internet (0.0.0.0/0). This default configuration exposes provisioned instances to potential unauthorized access and brute-force attempts.

terraform-gcp