The Agent Skills Directory

[PROMPT_INJECTION]: The scripts/analyze.sh script is vulnerable to indirect prompt injection through schema confusion.
Ingestion points: The --image-url and --query arguments in scripts/analyze.sh accept untrusted external data.
Boundary markers: None; the script lacks delimiters or instructions to ignore embedded commands within the input strings.
Capability inventory: The script performs network operations using curl to interact with the fal.ai API.
Sanitization: Absent; inputs are directly interpolated into a JSON payload string (e.g., PAYLOAD="{\"image_url\": \"$IMAGE_URL\"}"), which allows an attacker to manipulate the JSON structure if the input contains unescaped double quotes.
[COMMAND_EXECUTION]: The skill executes a local bash script that utilizes system utilities such as curl and jq. Additionally, the script's --add-fal-key functionality writes sensitive API keys to a local .env file in plaintext, which is a suboptimal security practice for credential management.

fal-vision