When deploying traditional software, you need firewalls and monitoring for solid cybersecurity. But when you’re deploying AI-powered applications, how are you protecting them? Cygnal is Gray Swan's answer: a powerful AI safety layer to protect your AI endpoints. Beyond cybersecurity, Cygnal is AI security.
Cygnal is a state-of-the-art solution designed to secure any AI model through advanced input filtering, output filtering, and continuous monitoring to ensure security and adherence to your policies. It supports high-performance streaming, multimodal inputs and outputs, customizable policies, and real-time adaptation to violation attempts. With a simple one-line code change, you can secure any AI endpoint, letting your business deploy AI-powered applications and workflows with confidence.
The training techniques behind Cygnal's unparalleled performance come from Gray Swan's world-leading research team, which provides adversarial robustness products and analysis to labs like OpenAI, Anthropic, xAI, and the UK AISI. Join us at the forefront of AI security by wrapping one of your AI endpoints today.
AI is fundamentally different from current software systems. While all large-scale software has the potential for security vulnerabilities, AI systems raise this risk substantially. Whereas traditional software follows clear logical rules specified by programmers, modern AI systems respond to developer and user commands in unexpected and potentially harmful ways. Cygnal is designed to give developers substantially improved degrees of safety and control over the models they deploy.
At Gray Swan, safety and security always come first. We subject Cygnal to rigorous, ever-evolving safety and security testing that is continually being trained and adapted to ensure that it remains as resilient as possible under pressure. As new threats and attack vectors arise in the wild, Cygnal learns and fortifies itself against them, leading to ever-increasing robustness and adherence to your policies.
Cygnal is designed to deliver blazing fast performance with minimal latency overhead. We work hard to optimize not just literal but also perceived latency, especially for streaming use cases, so that you can take advantage of all the raw speed of your underlying model. Whether you need natural language understanding, multimodal inputs and outputs, or the ability to call functions, Cygnal delivers with precision and efficiency. As a text moderation endpoint, Cygnal outperforms other leading solutions while offering superior robustness.
For organizations that need to keep their AI close by, Cygnal can be deployed on-premises under an enterprise agreement. This option provides the ultimate in security and customization, allowing you to harness the future of secure AI while keeping your data within your own infrastructure. Contact us to learn more!
Accessing Cygnal is straightforward and flexible. Cygnal understands all popular formats for calling AI models, which means all you have to do is change one URL in your code, and Cygnal will adapt to the target model’s API. Want to get started with just testing on a small fraction of traffic? Easy–you can configure everything with simple URL parameters, or in the Gray Swan dashboard. This compatibility ensures seamless integration into your existing workflows and applications, allowing you to leverage Cygnal’s power with ease.
Cygnal offers the latest in safe and secure AI while also featuring leading performance on blocking harmful requests and allowing permitted requests. A demonstration of our commitment to deliver capabilities without compromising on robustness, it strikes an unprecedented balance between the two.
Cygnal delivers an industry-leading attack block rate of 99.98% in a double-blind public evaluation of OpenAI models from hundreds of thousands of red teaming attempts, turning the least-robust model into by far the most-robust model with only a 0.02% attack success rate (ASR).
Cygnal’s performance on the standard Toxic-Chat benchmark of challenging harmful chat requests is an industry-leading 82.3% AUPR (area under precision-recall curve), balancing exceptional precision in detecting harmful content with comprehensive coverage to catch what others miss.
Moderation is worthless without precision. The training techniques that power Cygnal’s filtering result in a best-in-class 97.4% AUPR score on the difficult XSTEST over-refusal benchmark. Cygnal minimizes false positives and doesn’t interfere with on-policy AI usage.
Contact [email protected] to explore Cygnal solutions for your organization.
Best-in-class performance with unparalleled security and safety. Deploy with confidence without sacrificing intelligence.
Harness the latest tools and results in adversarial AI to understand how your AI will stand up under the toughest conditions.
Staying safe and secure in the AI era requires staying ahead of the changing threat landscape.
Keep up to date on all things Gray Swan AI and AI Security.
