GitHub hosts HunyuanVideo, Tencent's open-source framework for large-scale video generation models, enabling AI-driven video creation.
Sunglasses
About Sunglasses
Sunglasses is a local-first security scanner designed to protect AI agents from prompt injection, tool poisoning, credential exfiltration, and other input-based attacks. It operates as a pre-filter, scanning prompts, files, repositories, web pages, and MCP tool metadata before any content reaches the model or CI pipeline. The tool enforces security policies by blocking malicious instructions, secret material, or poisoned content, with enforcement capabilities available for specific agent frameworks like Claude Code. Sunglasses uses a pattern database of 1,407 detections across 117 categories, achieving a median scan time of 0.26ms without relying on cloud calls or API keys. It is MIT-licensed and can be installed via pip, offering both a scanner and an enforcing firewall mode. The project provides extensive documentation, integration guides, and reference materials covering AI agent security fundamentals and compliance mappings.
Key features
- Scans prompts, files, repositories, web pages, and MCP tool metadata
- 1,407 detection patterns with daily updates
- 0.26ms median scan time with local execution
- Enforcing firewall for blocking secret exfiltration and policy violations
- GitHub Action for CI pipeline integration
- CLI for terminal-based scanning
- Pattern database browser for attack family categorization
- MCP Server mode for Claude & MCP hosts
Use cases
- Protecting AI agents from prompt injection attacks in production environments
- Securing CI pipelines by scanning pull requests for malicious content
- Hardening AI agent deployments against credential theft and tool poisoning
Pros
- Local-first operation with no cloud calls or API keys required
- 1,407 detection patterns across 117 categories for comprehensive threat coverage
- Zero-config installation via pip with sub-millisecond median scan times
- Enforcing firewall mode for blocking malicious tool calls and enforcing policies
- Open-source MIT license with active development and daily pattern updates
Cons
- Limited to input filtering; does not address model-side vulnerabilities
- Enforcing firewall currently supports specific agent frameworks like Claude Code
- No native support for non-text inputs beyond basic image scanning in the live demo
Frequently asked questions about Sunglasses
What is Sunglasses and what does it do?
Sunglasses is an open-source input firewall for AI agents that scans prompts, files, repositories, web pages, and MCP tool metadata for security threats such as prompt injection, tool poisoning, and credential exfiltration before any content reaches the model or CI pipeline.
Who should use Sunglasses?
Sunglasses is designed for developers and organizations deploying AI agents who need to protect against input-based attacks, ensuring secure operation in local or CI environments without relying on cloud calls or API keys.
How does Sunglasses work?
It operates as a pre-filter using a pattern database of 1,407 detections across 117 categories, scanning content in real-time with a median scan time of 0.26ms, and can enforce security policies by blocking malicious instructions or secret material.
Does Sunglasses require an internet connection or API keys?
No, Sunglasses is a local-first tool that performs all scans and enforcement without cloud calls or API keys, ensuring privacy and offline operation.
What frameworks or tools does Sunglasses integrate with?
Sunglasses integrates with multiple agent frameworks including Claude Code, LangChain, CrewAI, OpenAI Agents, and MCP hosts, with support for GitHub Actions and CLI scanning.
How can I get started with Sunglasses?
You can install Sunglasses via pip with a single command, and it can be set up to run in under a minute, with options for scanner or enforcing firewall modes depending on your needs.
Sunglasses Website Engagement
Last Update: 24 days ago