The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted text and file content by reading it into memory and sending it to an external API for analysis. This creates a potential surface for indirect prompt injection if the processed content contains instructions that attempt to influence the agent's interpretation of the detector's results.
Ingestion points: Content is ingested from local text files via shell commands (curl -F "files=@path") and Node.js scripts (readFile(process.argv[2])).
Boundary markers: The skill uses structured JSON objects to wrap data sent to the API, which helps delineate data from instructions.
Capability inventory: The skill utilizes file system read access and network communication via curl and the Node.js fetch API.
Sanitization: The implementation uses JSON.stringify and jq -Rs to sanitize input and prevent malformed request bodies, adhering to standard security practices.
[COMMAND_EXECUTION]: The documentation provides multiple shell command examples using curl and jq to interact with the API and process local files. These commands are intended for manual execution by the user or agent during integration tasks.

gptzero