The Agent Skills Directory

PROMPT_INJECTION (HIGH): The slide generation script scripts/generate_slide_image.py is vulnerable to Indirect Prompt Injection. It ingests natural language prompts and image attachments from untrusted sources and passes them to an AI generation process without any sanitization or protective boundary markers.\n
Ingestion points: The prompt CLI argument and --attach image files in scripts/generate_slide_image.py.\n
Boundary markers: Absent. There are no instructions or delimiters used to separate user input from system instructions.\n
Capability inventory: File writing (output slide images) and execution of an external AI generation script via subprocess.run.\n
Sanitization: None. Raw strings and files are passed directly to the execution pipeline.\n- COMMAND_EXECUTION (MEDIUM): scripts/generate_slide_image.py performs dynamic execution of a companion script (generate_slide_image_ai.py) using subprocess.run. While it uses a list of arguments to mitigate shell injection, the execution flow depends on the presence and integrity of a file that is not included in the provided source code.\n- PROMPT_INJECTION (MEDIUM): scripts/pdf_to_images.py processes external PDF files, which serves as an indirect injection vector. Malicious content within a PDF could be converted into images and subsequently interpreted by an AI agent as valid instructions, potentially leading to unauthorized actions during the presentation workflow.\n
Ingestion points: User-provided PDF file paths in scripts/pdf_to_images.py.\n
Capability inventory: Full file system read access for PDFs and write access for image generation.

scientific-slides