Pular para o conteúdo
TakoAPI
Voltar às skills

safe-web

por adamnaghs

Web & Frontend Development
1.2kDownloads informados pelo clawskills.sh
1Estrelas
0Visualizações
0 curtidaEntre para curtir esta skill
—Ainda sem avaliações
Sua avaliação

Resumo

Secure web fetch and search with **PromptGuard** scanning.

Descrição

O que esta skill faz

Wraps web fetch and search operations with PromptGuard scanning to detect prompt injection attacks hidden in web content. Any page or search result that fails the scan is blocked before it enters the AI context window. Operates fail-closed: if scanning cannot run, the tool returns an error rather than unverified content.

Native web fetch tools return content without injection scanning, so malicious pages can silently alter AI behavior without detection.

Quando usá-la

  • Fetching articles from unfamiliar or untrusted URLs
  • Scraping competitor or news sites without exposing the AI to injected instructions
  • Searching the web during automated research pipelines
  • Processing user-supplied URLs in an AI agent workflow
  • Auditing web pages suspected of containing embedded injection attempts

Exemplo de fluxo de trabalho

ENTRADA

User asks: Summarize this article: https://unknown-blog.com/post

AGENTE
  1. 1

    Runs 'safe-web fetch https://unknown-blog.com/post'

  2. 2

    Downloads the page and strips scripts and styles with BeautifulSoup

  3. 3

    Scans extracted text with PromptGuard

  4. 4

    Blocks content and returns a SHIELD report if injection patterns are detected

  5. 5

    Returns clean text to the AI if the scan passes

SAÍDA

Clean article text ready for summarization, or a security alert listing matched injection patterns and the blocked URL

Requisitos

  • PromptGuard installed from clawhub.ai/seojoonkim/prompt-guard
  • Brave Search API key set as BRAVE_API_KEY environment variable (required for search command, free tier available at brave.com/search/api/)

Instalação

clawhub install adamnaghs/safe-web

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

Links