The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted user data, such as text prompts and media URLs, which are interpolated into shell commands and API request bodies. This creates a surface for indirect prompt injection or command injection if the input is not properly handled by the execution environment.
Ingestion points: Found in all command files where $ARGUMENTS or user-supplied URLs are used, such as commands/embed.md, commands/generate-image.md, and commands/transcribe.md.
Boundary markers: The instructions lack delimiters or specific directives for the agent to treat user input purely as data, which could lead the agent to follow instructions embedded within the input.
Capability inventory: The skill uses curl for network requests and interacts with the local file system by writing temporary files to /tmp (e.g., in commands/ocr.md).
Sanitization: There is no explicit instruction for shell-escaping or strict validation of user-supplied strings before they are passed to the curl command line.
[COMMAND_EXECUTION]: The skill relies heavily on shell-based curl commands to interact with the deAPI backend. User-supplied variables are placed directly into the command string.
Evidence: commands/ocr.md and commands/remove-bg.md execute curl -s -o /tmp/... "{image_url}" where {image_url} is derived from user input.
Evidence: commands/embed.md uses curl -s ... -d '{"input": "$ARGUMENTS", ...}' which places raw user text into a shell command context.

deapi