Detects text-input threats from untrusted sources and outputs severity results.
npx skills add https://github.com/dgriffin831/clawhub-skills --skill input-guard
Please help me install this Agent Skill. Skill: input-guard Source: https://github.com/dgriffin831/clawhub-skills/tree/main/input-guard Command: npx skills add https://github.com/dgriffin831/clawhub-skills --skill input-guard
Input Guard protects AI agents by scanning text fetched from untrusted external sources for prompt injection attempts, returning severity, findings, and alerts before the content is ever processed.
Use the Input Guard to scan external text before it is fed to the agent:
High-intent search queries and answers about installing and using this skill.
To detect prompt injection in external content, you can scan text using pattern-based detection across 16 threat categories before your AI agent processes it. This covers sources like web pages, tweets, search results, and API responses.
A pattern-based security scanner identifies 16 threat categories including instruction override, role manipulation, system mimicry, jailbreak, and data exfiltration. It supports multi-language detection for English, Korean, Japanese, and Chinese prompts.
To scan fetched text for prompt injection, run the scan script via bash with direct text, a file path, or stdin. You can also execute the Python script directly with the --json flag to receive structured severity results.
Yes, you can use optional LLM analysis to catch evasive prompt injection attacks by configuring OpenAI or Anthropic. This feature uses a shipped MoltThreats taxonomy to identify subtle threats that pattern-based detection might miss.
The base prompt injection scanner requires zero external dependencies, using only the Python standard library. Optional LLM-powered analysis and MoltThreats reporting require the requests library and an OpenAI or Anthropic API.
To receive alerts for high risk prompt injection attempts, integrate the scanner with OpenClaw alert channels. It automatically notifies humans about MEDIUM+ and HIGH risk findings detected in the external content.
Driving-TeacherFetch and parse tweet text, author, and engagement metrics from Twitter/X URLs.
sncollectiveConfigure tusd v2 servers for resumable uploads with S3 storage and HTTP hooks.
allthingslinuxRoute alerts and coordinate incidents across PagerDuty, Alertmanager, and Slack.
cmaennerTest LLM applications against prompt injection techniques using the Arcanum PI Taxonomy.
