Introducing Fiddler Guardrails: Safeguarding LLM Applications from Safety and Security Risks

As a pioneer in AI Observability, our mission has always been to help enterprises confidently launch AI systems into production — whether LLM or ML applications — ensuring they are performant, safe, and trustworthy. By confidently moving AI models and applications beyond experimental and evaluation phases, enterprises can accelerate ROI while maintaining trust, safety, and transparency in their AI initiatives. 

Fiddler Trust Service: The Foundation for LLM Observability

Last year, we introduced the Fiddler Trust Service, an extension of our Fiddler AI Observability and Security platform. Powered by Fiddler Trust Models, the Fiddler Trust Service enables teams across application development, LLM engineering, and security to score LLM prompts and responses across dozens of trust dimensions like hallucinations, toxicity, PII leakage, prompt injection attacks, jailbreaking attempts, and more. We've continued to enhance our Trust Models: 

  • Fast: With a <150ms latency, the models are optimized for rapid scoring and monitoring, ensuring enterprises can rapidly detect and resolve LLM issues.
  • Scalable: Handles enterprise traffic volumes out of the box.
  • Cost-Effective: Trust Models are task-specific, designed to optimize performance and minimize computational overhead.
  • Secure: Built with enterprise-grade security standards to protect sensitive data.

Fiddler Guardrails: Moderating LLM Applications in Runtime

Building on this foundation, we’re thrilled to launch Fiddler Guardrails, a powerful solution designed to protect enterprises and their users from harmful and costly LLM risks. In live environments, Fiddler Guardrails moderates risky issues in prompts and responses, such as hallucinations, safety violations, prompt injection attacks, and jailbreaking attempts. Guardrails can be deployed in VPC and air-gapped environments, enabling enterprises to maintain compliance and protect sensitive data in even the most regulated industries.

Fiddler AI Observability and Security for LLMOps Fiddler Trust Service for LLM scoring and monitoring

How Fiddler Guardrails Work 

Fiddler Guardrails extend the capabilities of the Fiddler Trust Service, enabling enterprises to score and moderate risky prompts and responses rapidly. With a robust framework for moderating and enforcing accuracy, safety, and trust standards, Fiddler Guardrails operate at enterprise scale out of the box, handling over 5 million requests per day with latency under 150ms.

Track total number of Jailbreak attempts detected and blocked.
Fiddler Guardrails flags hallucination, and safety issues in LLM conversations

Leveraging Fiddler Trust Models, Guardrails moderates by scoring key hallucination and safety metrics such as:

Hallucination Metrics:

  • Faithfulness / Groundedness

Safety Metrics:

  • Illegal
  • Hateful
  • Harassing
  • Racist
  • Sexist
  • Violent
  • Sexual
  • Harmful
  • Unethical
  • Jailbreaking

These LLM metrics are computed via Fiddler’s secure inference layer, enabling rapid, and reliable scoring and analysis with minimal latency. This inference layer eliminates the need for routing sensitive data through external platforms, ensuring all that data stays secure within the production environment. 

Security and AI teams can build and customize Guardrails by calling an API to their Fiddler platform to moderate prompts and responses. With just three lines of code, they can customize Guardrails to their organization’s risk standards, and set threshold scores to accept or reject prompts and responses. 

Risky prompts and responses are monitored over time, providing actionable insights through dashboards and reports. Security and AI teams can conduct root cause analysis and model diagnostics to address LLM risks effectively.

3-Step AI Observability and Security for your LLM Applications

Fiddler AI Observability and Security for LLMs showcasing three steps: Trust Service for LLM Scoring and Guardrails, LLM Monitoring, and LLM Analytics.

Fiddler Guardrails is essential to safeguard LLM applications in production. 

  1. Score: By enabling Security and AI teams to score and customize guardrails tailored to their specific risk preferences, Guardrails provides the first line of defense and ongoing oversight. 
  2. Monitor: Critical LLM metrics can be monitored over time through custom reports.
  3. Analyze: Conduct root cause analysis and diagnostics for actionable insights. 

Watch how Fiddler Guardrails can protect your LLM applications in production.