João Henrique
b541f502ba
- Adicionado estrutura completa do projeto - Configurado MCP server para Premiere Pro - Adicionado documentação e skills - Configurado Gitignore para o projeto
21 lines
1.1 KiB
Markdown
Executable File
21 lines
1.1 KiB
Markdown
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.
|
|
|
|
- [MCP prompts security](https://modelcontextprotocol.io/specification/2026-07-28/server/prompts)
|
|
|
|
## 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.
|