The Agent Skills Directory

[EXTERNAL_DOWNLOADS] (LOW): The script generate_image.sh uses curl to communicate with api.openai.com. While this is the intended functionality of the skill, it involves network requests to a non-whitelisted domain according to the security framework.
[COMMAND_EXECUTION] (SAFE): The skill executes shell scripts and Bun commands. Analysis of generate_image.sh shows that input variables (like --prompt) are handled using jq --arg, which prevents shell command injection. Arguments are generally well-quoted throughout the script.
[DATA_EXFILTRATION] (LOW): The script allows the user/agent to specify an --input-image path. This file is then uploaded to OpenAI's API via curl -F. While intended for image editing, a malicious actor could use indirect prompt injection to trick the agent into uploading sensitive text files (e.g., .env or SSH keys) by passing them to the --input-image parameter.
[INDIRECT_PROMPT_INJECTION] (LOW): The skill exhibits an attack surface for indirect prompt injection.
Ingestion points: The --input-image and --mask parameters in generate_image.sh allow reading any file accessible to the agent process.
Boundary markers: No markers or validation steps are present to ensure the files provided are actually images before transmission.
Capability inventory: The skill can read local files (curl -F "image=@...") and perform network operations to an external API.
Sanitization: There is no sanitization or file-type validation (e.g., checking magic bytes) on the input files.

openai-image-generation