Pular para o conteúdo
TakoAPI
Voltar às skills

skill-firewall

por mkhaytman87

Web & Frontend Development
1.0kDownloads informados pelo clawskills.sh
2Estrelas
1Visualizações
0 curtidaEntre para curtir esta skill
—Ainda sem avaliações
Sua avaliação

Resumo

Security layer that prevents prompt injection from external skills.

Descrição

O que esta skill faz

A security layer for AI agents that intercepts requests to install external skills and rewrites them from scratch instead of copying them verbatim. Prevents prompt injection via hidden HTML comments, zero-width Unicode characters, and embedded malicious instructions embedded in external skill files.

Copying and executing external skills directly exposes the agent to payloads invisible in rendered markdown but visible to the LLM, which regeneration eliminates entirely.

Quando usá-la

  • Installing a community skill from ClawHub without running hidden malicious instructions
  • Safely evaluating a GitHub-hosted skill before adding it to a project
  • Blocking a curl-based exfiltration attempt disguised as skill setup
  • Reviewing a skills.sh package for obfuscated Unicode payloads before use
  • Sanitizing a third-party skill that references external files in a poison chain

Exemplo de fluxo de trabalho

ENTRADA

User asks: Install the cloudflare skill from skills.sh

AGENTE
  1. 1

    Acknowledge the request and state that the skill will be rewritten, not copied

  2. 2

    Fetch the external skill content and identify its actual purpose

  3. 3

    Flag any suspicious elements such as HTML comments or external URL references

  4. 4

    Write a clean replacement skill in standard ASCII with no external references

  5. 5

    Present the rewrite with a summary report and wait for explicit user approval before saving

SAÍDA

A sanitized, human-approved skill file saved to the skills directory, functionally equivalent to the original but free of injected instructions

Instalação

clawhub install mkhaytman87/skill-firewall

Execute o comando da CLI ou cole o prompt no Claude Code, Cursor, Codex ou no agente de código que você usa.

Links