Sunglasses: An open-source input scanner for AI agents (github.com)

SUNGLASSES is a newly launched open-source input scanner designed to enhance the security of AI agents by serving as a local filtering layer that scans and cleans input before it reaches the agent. This tool is significant for the AI/ML community because it addresses the growing threat of sophisticated attacks that often disguise malicious instructions within seemingly benign inputs such as emails, web content, images, and more. By operating entirely offline and locally, SUNGLASSES ensures that sensitive data never leaves the user’s device, enhancing privacy and security without the need for cloud connections or API keys. The scanner boasts a wide range of capabilities, including the ability to detect prompt injections in 23 languages, credential exfiltration, command injections, and various forms of social engineering. With advanced features like OCR for image text, audio transcription, and extensive compatibility with various file types, SUNGLASSES processes data swiftly with an average response time of less than 1 millisecond for text. The tool operates under a robust mechanism composed of multiple normalization techniques and pattern recognition rules, providing a high recall rate while maintaining a low false-positive rate. SUNGLASSES can be used independently or integrated with other security platforms, marking an important step in the ongoing efforts to safeguard AI agents against evolving threats.