The Agent Skills Directory

[DYNAMIC_EXECUTION]: The importDataset function constructs a Python command string using direct interpolation of the sourceUrl parameter. Because this string is executed via python3 -c within a generated Kubernetes Job, a malicious URL could be used to inject and execute arbitrary Python code inside the container environment.\n- [COMMAND_EXECUTION]: The exportCheckpoint function generates a shell command for data transfer by concatenating the destBucket variable into a gsutil command string. The lack of shell escaping or validation for metacharacters allows for potential command injection if the bucket name is manipulated.\n- [INDIRECT_PROMPT_INJECTION]: The skill represents an attack surface for indirect prompt injection by processing untrusted external inputs and using them to define executable logic and Kubernetes resources.\n
Ingestion points: External data enters through the sourceUrl parameter in importDataset and the destBucket parameter in exportCheckpoint.\n
Boundary markers: None; external inputs are directly embedded into executable strings without delimiters or escaping.\n
Capability inventory: The skill uses the @kubernetes/client-node library to create and manage batch jobs and generates shell commands for data synchronization.\n
Sanitization: The exportCheckpoint function includes a basic region-matching check but does not filter shell metacharacters. The importDataset function performs no validation or sanitization on the source URL.

coreweave-data-handling