The Agent Skills Directory

[DATA_EXFILTRATION]: The skill instructs the agent to transmit the "complete, verbatim contents" of local configuration files and project data to the ask_cloud_operator tool on the autocloud MCP server. Although it contains explicit instructions to redact secrets and avoid credential files, the automated transmission of potentially sensitive project source code to a remote agent constitutes a data exposure risk.\n- [INDIRECT_PROMPT_INJECTION]: The skill processes user queries and local file contents (Terraform, YAML) "as-is" and passes them to a secondary agent, creating an attack surface for instructions embedded in analyzed data.\n
Ingestion points: Local configuration files (*.tf, *.yaml, etc.), diffs, error logs, and user queries (SKILL.md).\n
Boundary markers: External data is delimited using headings containing the file paths (SKILL.md).\n
Capability inventory: The skill can read local filesystem data, write generated code to the workspace, and invoke the ask_cloud_operator MCP tool (SKILL.md).\n
Sanitization: Implements a rule to redact passwords, tokens, and keys, and explicitly forbids sending files like .env, *.tfstate, and kubeconfig (SKILL.md).\n- [COMMAND_EXECUTION]: The skill utilizes gcloud config get-value project as a fallback mechanism to identify the active GCP project context (SKILL.md).

operator