The Agent Skills Directory

[PROMPT_INJECTION]: The skill presents a surface for indirect prompt injection through the processing of external configuration data.\n
Ingestion points: The agent reads and reviews assets/example_config.yaml and assets/pipeline_template.yaml. Additionally, scripts/init_pipeline.sh accepts external JSON configuration files via the --config parameter.\n
Boundary markers: There are no explicit delimiters or instructions provided to the agent to distinguish between configuration data and instructions, which may lead to the agent following malicious commands embedded in these files.\n
Capability inventory: The skill allows the use of high-impact tools including Bash(git:*), Bash(docker:*), and Bash(kubectl:*), as well as Write and Edit operations.\n
Sanitization: No evidence of validation, escaping, or content filtering is present in the provided templates or scripts.\n- [COMMAND_EXECUTION]: The skill requests access to powerful CLI tools and performs file system operations without sufficient input validation.\n
Evidence: The allowed-tools section in SKILL.md enables git, docker, and kubectl. These tools provide a broad attack surface if the agent is influenced by malicious input.\n
Evidence: The scripts/init_pipeline.sh script creates a directory structure using a user-provided project_name without validating for path traversal characters (e.g., ../). This could allow the script to create directories or files in unintended locations on the host system.

orchestrating-deployment-pipelines