sunglasses

Sunglasses is an open-source input firewall designed specifically for AI agents, aiming to protect them from a range of security threats before malicious content reaches the model or enters a CI workflow. The tool focuses on scanning and filtering prompts, files, repositories, web pages, and metadata associated with tools and multi-agent control protocols (MCP) to detect and block attacks such as prompt injection, tool poisoning, credential exfiltration, encoding smuggling, and memory poisoning. With a database of 1,089 detection patterns across 65 categories, Sunglasses is built to identify a broad spectrum of adversarial input techniques, including credential theft, command injection, data exfiltration, social engineering, and various encoding-based evasions.

26 milliseconds. It is distributed as a Python package, installable via pip, and can be integrated into AI agent workflows with minimal setup. Sunglasses provides several usage options: a CLI for scanning files and pipelines from the terminal, a GitHub Action for scanning pull requests in CI environments, and direct integration into agent code using its detection library. There is also a live demo available in the browser for testing text, GitHub repositories, and images without installation. Pattern updates are delivered daily, ensuring ongoing protection against emerging threats.

The tool is compatible with a variety of AI agent frameworks and stacks, including Claude Code, LangChain, CrewAI, OpenAI Agents SDK, LlamaIndex, and others, as explicitly listed in its documentation. Sunglasses emphasizes compliance with security standards such as the OWASP LLM Top 10, OWASP Agentic Top 10, and MITRE ATLAS, and provides reference materials like the Encyclopedia of AI Agent Security and the MCP Attack Atlas. Licensing is under the MIT license, and the package is free to use. Its primary audience includes developers and operators of AI agents who require robust, local-first security filtering to guard against input-based attacks.