The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The skill downloads and executes the Google Cloud SDK installation script via a shell pipe from an official Google domain.
[EXTERNAL_DOWNLOADS]: Fetches the AWS CLI installation package and various security assessment tools (e.g., ScoutSuite, Pacu) from official and well-known public registries.
[COMMAND_EXECUTION]: Provides instructions to execute high-impact commands on target environments, such as running scripts on virtual machines via Invoke-AzVMRunCommand and creating administrative users with az ad user create.
[PERSISTENCE]: Contains detailed procedures for establishing long-term access to target cloud infrastructures, including the creation of backdoor service principals in Azure and new access keys in AWS.
[INDIRECT_PROMPT_INJECTION]: The skill represents an attack surface for indirect prompt injection because it ingests untrusted data from external sources (e.g., cloud resource enumeration, S3 bucket listings, metadata service responses) and processes this data while possessing the capability to modify environment configurations and execute arbitrary commands.
Ingestion points: External enumeration tool outputs (cloud_enum.py, ip2provider.py), cloud resource descriptions (Get-AzResource, aws s3 ls), and metadata service responses (http://169.254.169.254/).
Boundary markers: None provided in the instructions to delimit untrusted data or warn the agent against embedded instructions.
Capability inventory: Subprocess execution for CLI tools, Invoke-AzVMRunCommand for VM script execution, and administrative account/credential management.
Sanitization: No explicit sanitization or validation of data retrieved from cloud environments before it is processed by the agent.

Cloud Penetration Testing