The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The file upload demonstration in EXAMPLES.md contains a directory traversal vulnerability due to unsafe handling of user-supplied filenames.
Ingestion points: The upload_file and upload_multiple_files endpoints in EXAMPLES.md accept UploadFile objects containing user-controlled filenames.
Boundary markers: The instructions do not include boundary markers or warnings advising the user to sanitize or ignore instructions embedded within external file metadata.
Capability inventory: The skill includes file system write capabilities via shutil.copyfileobj and directory operations in EXAMPLES.md.
Sanitization: The code lacks sanitization for the file.filename attribute; it directly concatenates this untrusted string to the UPLOAD_DIR path (UPLOAD_DIR / file.filename). This allows a malicious actor to provide a filename containing traversal sequences (e.g., ../../etc/passwd) to write files to arbitrary locations on the server.

fastapi-development