The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill documents architectures and setups for parsing unstructured user input via Telegram chats (text, documents, photo, voice). These entry points are integrated into LLM prompts (e.g., n8n OpenAI nodes and conversational features), which introduces a pipeline risk where malicious user input can control bot behavior.
Ingestion points:
File: SKILL.md (e.g., text, photo, document, voice inputs mapped in workflows/triggers)
Boundary markers:
Absent. Input values like {{ $json.message.text }} are interpolated directly into prompts without protective delimiters or instruction isolation patterns.
Capability inventory:
File: SKILL.md (e.g., capabilities include telegram_send_message, telegram_send_photo, vision processing, and automation integrations like web_search or exporting results to spreadsheets)
Sanitization:
Absent. Content is processed directly without filtering/escaping mechanisms.

telegram-bot