The Agent Skills Directory

PROMPT_INJECTION (LOW): The skill possesses a significant attack surface for indirect prompt injection. It ingests untrusted data from multiple external sources and interpolates it directly into LLM prompts without sanitization or boundary markers.
Ingestion points: Data is ingested via extract_github_content, extract_web_content, extract_youtube_content, and extract_email_content in examples/document_processors.py.
Boundary markers: Absent. The generate_follow_up_questions function in examples/conversation_manager.py interpolates the first 500 characters of the context and response directly into a prompt template (Context: {context[:500]}) without using delimiters or instructions to ignore embedded commands.
Capability inventory: The skill performs network operations (requests to GitHub and web URLs) and file system reads (mailbox paths).
Sanitization: Absent. The preprocess_document function in examples/text_processor.py cleans whitespace and special characters but does not filter for malicious instructional content or escape characters that might break prompt structures.

document-chat-interface