Agent Guard — Prompt Injection Protection

Trust Score 80/100

Realtime pattern-based detection and sanitization for prompt and command injection, designed to screen GitHub issues, sanitize inputs, and stop dangerous comman

triggers:prompt injectioncommand injectionsanitizeagent securitygithub issueclinejectionsanitize input

GitHub SKILL.md

What it does

Agent Guard is a real-time security framework that detects and sanitizes prompt and command injection attempts aimed at AI agents. It applies pattern-based detectors for command execution (npm, curl|bash, rm -rf), prompt-injection phrases (ignore previous instructions, developer mode), social engineering cues, and specialized GitHub issue screening for Clinejection-style attacks. The tool can replace dangerous segments with safe placeholders and produce threat reports.

When to use it

Use Agent Guard wherever untrusted text could influence agent behavior: screening GitHub issue titles/bodies, sanitizing user-supplied prompts, or pre-filtering web-scraped content before model consumption. It is appropriate in CI pipelines, moderation layers, or any agent pipeline that executes commands or generates actions based on external text.

What's included

Scripts: core Python modules including agent_guard.py, mcp_server.py, and cli.py for local testing and integration.
References: pattern lists and configuration samples for thresholds and caching.
Instructions: installation notes for using as an OpenClaw skill or Claude MCP server, example CLI usage, and config.json for tuning thresholds.

Compatible agents

Built primarily for OpenClaw/Claude MCP deployments but usable as a standalone Python library. It integrates easily with agent runtimes that can call a local sanitize/analyze endpoint or run a CLI check before executing actions.

Quality assessment: substantive, actionable, and security-focused — suitable to save.

Audit Summary

Agent Guard provides pattern-based prompt injection detection and sanitization for AI agents, inspired by the Clinejection incident. The SKILL.md is detailed and well-structured with clear usage examples, API docs, and detection pattern lists. No scripts were bundled for execution testing. The skill claims zero dependencies for core engine (Python stdlib only), which is good for security but unverified without runnable scripts.

Watch Out

Claims zero dependencies but no scripts available to verify
Performance claims (50K analyses/sec, 98.7% detection) are unsubstantiated without test suite
Social engineering detection patterns may produce false positives on legitimate urgent content

Notes

Security-focused skill — no dangerous commands, no exfiltration, no hardcoded credentials. Minor concern: the SKILL.md references pip install and curl patterns as detection targets, not as commands to execute. Well-documented with clear frontmatter. No bundled scripts so static analysis only.

Information

Repository: agentguard
Stars: 101

Trust Score

Overall80

Security92

Code Quality68

Architecture70

Usefulness72

Related Skills

Readwise Reader Document Management

Manage Readwise Reader documents: list, save, search, move, tag, highlight, export and bulk-edit via official and custom CLIs.

Bounty Hunter — Atlas

Persona skill: 'Atlas' — a profit-focused developer persona for discovering, evaluating and executing paid bounties or freelance tasks with ROI-aware workflows.

TradeOS

Natural-language driven CEX trading and portfolio management for agents: API key vaulting, multi-exchange orders, DCA, arbitrage scanning, alerts and security r

ezBookkeeping API Tools

Command-line API tools for ezBookkeeping: record and query transactions, retrieve accounts/categories/tags, and fetch exchange rates for self-hosted personal fi

Kickbox Automation (Composio / Rube MCP)

Automate Kickbox workflows via the Composio Kickbox toolkit on a Rube MCP server: discover tools, manage connections, and execute schema-compliant tool calls.

Beaconstac Automation

Automate Beaconstac QR code and link management using the Rube MCP (Composio) toolkit.

Jx Component Library

Build server-rendered UIs using reusable .jx components in Python/Jinja2 with typed props and automatic asset management.

Tmux Bridge

Give an agent controlled access to a local tmux terminal bridge so it can run shell commands, capture output, and manage sessions on the user's machine.

Back to Skills