The Agent Skills Directory

[Indirect Prompt Injection] (LOW): The AI/LLM streaming example in SKILL.md directly uses user input in a prompt, creating an injection surface.\n
Ingestion points: request.json() in SKILL.md (reading the prompt field).\n
Boundary markers: Absent. The input is directly passed as message content without delimiters or protective instructions.\n
Capability inventory: openai.chat.completions.create call in SKILL.md.\n
Sanitization: Absent. No validation or filtering is performed on the user-provided prompt.\n- [Data Exposure & Exfiltration] (LOW): The file upload example in references/http-methods.md is vulnerable to path traversal due to unsanitized filename usage.\n
Evidence: The code uses file.name directly in path.join(process.cwd(), 'public/uploads', filename) without removing traversal sequences like ...\n
Risk: An attacker could potentially write files to arbitrary locations outside the intended directory, affecting system integrity.

route-handlers