Gray Swan AI Launches Tools to Safeguard Enterprise Deployments from Malicious Use

Gray Swan AI, a pioneering company in the field of AI safety and security, has officially launched its platform with two key products aimed at helping enterprises mitigate risks associated with their AI systems. The company's mission is to provide robust tools that ensure AI components behave as intended, even in the presence of malicious users.

What Changed Technically?

The core issue Gray Swan addresses is the growing risk of AI models, such as large language models (LLMs), being used in ways that violate their intended behavior. These risks include generating harmful or inaccurate information, which can have severe consequences for enterprise deployments. To tackle this, Gray Swan has introduced two products: Gray Swan Shade and Gray Swan Cygnet.

Gray Swan Shade: Continuous AI Risk Assessment

• Comprehensive Evaluation: Shade is a tool designed to continuously assess the risks and vulnerabilities of AI components in real-time.
• Automated Safety Assessment: It leverages state-of-the-art automated techniques to identify potential threats that human red teams might miss.
• Ongoing Updates: The Gray Swan team continually updates Shade with the latest advancements in AI safety, ensuring developers have access to the most current risk assessment methods.

Gray Swan Cygnet: A Safe and Resilient LLM

• Engineered for Safety: Cygnet is a large language model (based on Llama-3) that has been specifically engineered and tuned to be highly resilient to powerful forms of attack.
• Key Capabilities Retained: Despite being a smaller model, it maintains the essential capabilities of its base model while offering enhanced security features.
• Free Experimentation: Users can experiment with Cygnet for free via Gray Swan's web interface. For bulk usage, per-token pricing is available through their API.

Why It Matters to Practitioners

The introduction of these tools addresses a critical gap in the current landscape of AI deployment. Developers and security teams now have access to:

• Reliable Risk Assessment: Continuous monitoring and evaluation of AI components to identify and mitigate risks.
• Enhanced Security: A safer LLM that can be integrated into existing systems without compromising performance.
• Scalability: Tools that can scale with the growing complexity of AI models and their applications.

Technical Details

• Shade:

  • Architecture: Shade integrates seamlessly into CI/CD pipelines, providing real-time risk assessments during development and deployment phases.
  • Benchmarks: Initial benchmarks show a significant reduction in false positives compared to traditional manual assessment methods.
  • Implementation: The tool uses a combination of static analysis, dynamic testing, and machine learning models to detect vulnerabilities.

• Cygnet:

  • Model Size: While smaller than some state-of-the-art LLMs, Cygnet has been optimized for safety without sacrificing key functionalities.
  • Research Backing: Recent studies by the Gray Swan team demonstrate that Cygnet is more resilient to adversarial attacks compared to other leading models.
  • Open Access: The model is available for free experimentation on the Gray Swan website, making it accessible for both researchers and developers.

Leadership and Research

Gray Swan was founded by leaders in AI safety and security. The team includes:

• Two faculty members from Carnegie Mellon University with extensive experience in AI safety research.
• Founders of the Center for AI Safety, a leading nonprofit organization dedicated to advancing the field of AI safety.

Their research has been widely recognized, featuring in major publications such as the New York Times, CNN, and the Washington Post.

Next Steps

For those interested in learning more about how Gray Swan's tools can enhance their AI deployments, you can contact the team at sales@grayswan.ai.