The Agent Skills Directory

[COMMAND_EXECUTION]: The invoke command in SKILL.md uses string interpolation for the {prompt} and {output} variables inside a shell command. If these variables contain shell metacharacters, it can lead to arbitrary code execution on the underlying host.
[CREDENTIALS_UNSAFE]: The load_api_key function in scripts/generate.py searches for .env files in parent directories up to three levels deep. The key name is constructed using string concatenation ("OPENROUTER_API" + "_KEY="), which is a technique used to evade static analysis tools searching for hardcoded secret names.
[DATA_EXFILTRATION]: The skill can read any local file provided to the --reference argument and send its contents to the OpenRouter API. This capability can be misused to exfiltrate sensitive local data.
[EXTERNAL_DOWNLOADS]: The skill connects to the external API at https://openrouter.ai to perform image generation tasks, which involves sending user data and prompts to a third-party service.
[PROMPT_INJECTION]: The skill processes untrusted input from prompts and batch JSON files without adequate sanitization or boundary enforcement.
Ingestion points: Prompts provided via the command line and batch files read from args.batch.
Boundary markers: None present; the prompts are directly embedded into the API request content.
Capability inventory: Network access through urllib.request and local file creation via open(output_path, "wb").
Sanitization: No validation or escaping is applied to the input data before it is sent to the AI model.

image-ai-generator