The Agent Skills Directory

[COMMAND_EXECUTION]: The skill contains a bash script (scripts/check-versions.sh) and Node.js templates that execute shell commands and system calls to manage dependencies and process data.
[EXTERNAL_DOWNLOADS]: Templates within the skill demonstrate how to fetch external resources, such as downloading generated images from OpenAI's servers via fetch.
[FILE_SYSTEM_ACCESS]: The templates/image-generation.ts file provides code to write generated image buffers to the local filesystem using fs.writeFileSync.
[INDIRECT_PROMPT_INJECTION]: The skill instructions and templates utilize OpenAI tools like file_search and web_search which ingest untrusted data from external files and websites into the model's context.
Ingestion points: External data is ingested through the outputs of the file_search (file content) and web_search (web search results) tools.
Boundary markers: Code templates do not implement specific delimiters or system instructions to prevent the model from following commands potentially embedded in the searched content.
Capability inventory: The skill environment has network access for API calls and filesystem access for saving processed results (e.g., images).
Sanitization: No explicit sanitization or filtering of the tool outputs is implemented in the provided examples.

openai-responses