Skip to main content
The model’s combined evaluation score is low, reflecting broad weaknesses across jailbreak, robustness, security, toxicity, and misuse assessments.

​
How to resolve

A low overall score indicates widespread weaknesses. To mitigate these risks, implement layered protections, such as input/output filters, anchoring responses in verified data (RAG), and restricting model-generated code to isolated environments that cannot access your internal systems especially when using tools. Prioritize your fixes based on how your application is used:
  • Customer-facing bots: Prioritize jailbreak and toxicity filters to protect your brand.
  • RAG-powered internal assistants: Focus on sensitive data (PII) detection and grounding responses in facts.
  • Code-generation APIs: Prioritize secure code execution and scanning for malicious technical patterns.
Refer to the specific remediation sections for the individual metrics (Jailbreak, Robustness, Toxicity, etc.) that contributed to this low score. If the score remains critically low even after adding these protections, reevaluate your choice of model, as it may be fundamentally unsuited for your safety requirements.

​
Risk

Governance/Compliance Low overall evaluation fails core expectations in the OWASP LLM Top-10, NIST AI RMF, and EU AI Act, all of which require broad safeguards across safety, robustness, and misuse prevention. A low score indicates the model cannot meet these end-to-end governance and safety requirements. Security A low overall score indicates systemic vulnerabilities and alignment deficiencies. This may result in:
  • Unsafe outputs
  • Harmful actions
  • Susceptibility to multiple attack vectors
  • Unreliability under real-world usage

​
Explanation

The overall score is a weighted aggregation across all evaluation categories. Low performance implies that the model’s safety posture is broadly inadequate, not just in isolated areas.

​
Specifications

Trigger: evaluation.overall_score < threshold. Severity:
  • CRITICAL: score < 0.2
  • HIGH: 0.2–0.35
  • MEDIUM: 0.35–0.5
  • LOW: 0.5–0.75
Applies to: Models