João Henrique
b541f502ba
- Adicionado estrutura completa do projeto - Configurado MCP server para Premiere Pro - Adicionado documentação e skills - Configurado Gitignore para o projeto
1.1 KiB
Executable File
1.1 KiB
Executable File
Recommendation 45: prompt and resource injection boundary
Evidence
The MCP prompt specification requires implementations to validate prompt inputs and outputs to prevent injection and unauthorized resource access. Premiere metadata and transcripts are untrusted project content.
Proposed improvement
Represent project-derived text as labeled data blocks with provenance, size limits, and escaping rather than concatenating it into trusted workflow instructions. Separate server-authored instructions, user arguments, and Premiere-derived content in every prompt renderer.
Acceptance criteria
- Adversarial clip names, markers, metadata, and transcripts cannot add server instructions.
- Resource links are independently authorized before rendering.
- Truncation preserves provenance and cannot splice delimiters ambiguously.
- A corpus tests injection, Unicode controls, nested markup, and oversized content.
Containment reduces instruction confusion but cannot guarantee model behavior.