The Agent Skills Directory

[COMMAND_EXECUTION]: The skill is designed to execute arbitrary shell commands and scripts through the type: shell attribute in workflow YAML files, which allows for high-privilege operations on the host system.\n- [EXTERNAL_DOWNLOADS]: The workflow system includes a dependencies feature that auto-installs external tools (e.g., tflint, checkov) based on specified versions, creating a path for untrusted software to enter the environment.\n- [EXTERNAL_DOWNLOADS]: Documentation examples illustrate the use of commands like wget to download external archives from remote URLs during the execution of a workflow step.\n- [REMOTE_CODE_EXECUTION]: Since the skill executes instructions directly from YAML configuration files, any compromise of the files in the stacks/workflows/ directory or the repository could lead to unauthorized remote code execution.\n- [PROMPT_INJECTION]:\n
Ingestion points: The agent loads and processes external workflow definitions from YAML files stored in the local file system.\n
Boundary markers: There are no documented mechanisms to sanitize command strings or distinguish between legitimate administrative instructions and malicious injections within the workflow files.\n
Capability inventory: The skill provides full shell access and infrastructure management capabilities (via Atmos/Terraform), which can be abused if the input YAML is manipulated.\n
Sanitization: The engine lacks a validation layer or whitelist for commands, relying entirely on the integrity of the configuration files.

atmos-workflows