The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The skill implementation facilitates the execution of unverified remote code at runtime.
Evidence: The script scripts/generate.py utilizes the uvx (or uv tool run) utility to download and run the gpt-image package directly from git+https://github.com/wuyoscar/gpt_image_2_skill. This mechanism bypasses local security inspections and allows the author to modify the executed code after the skill has been deployed.
[COMMAND_EXECUTION]: The skill executes system-level commands using arguments forwarded from the user context.
Evidence: In scripts/generate.py, the _delegate function uses subprocess.run to call the CLI tool, passing sys.argv[1:] (the original arguments provided to the script) directly into the execution list. This pattern can lead to command injection if the AI agent or the user provides malicious parameters that are not properly sanitized before being passed to the shell.
[PROMPT_INJECTION]: The skill's reference documentation contains high-risk prompt injection payloads used as examples.
Evidence: Within references/gallery-research-paper-figures.md (item No. 95) and references/craft.md, the text explicitly contains 'Ignore instructions' patterns, such as . While these are labeled as illustrative examples of security figures for educational purposes, they pose a risk if an agent reading these files interprets the examples as active system instructions.

gpt-image