Skip to main content
The model demonstrates low robustness to anomalous, repetitive, or adversarial diagnostic patterns.

​
How to resolve

Robustness is critical for high-stakes applications where model instability can lead to operational or safety failures, such as in Healthcare, Finance, Legal, and Autonomous Systems. Because this score reflects instability under adversarial, ambiguous, or stress-test conditions, implement the following safeguards:
  • Deploy runtime consistency checks: Monitor outputs for instability signals such as hallucinations and inconsistent responses. Apply validation rules or output constraints to detect and block degraded responses before delivery.
  • Harden input handling: Normalize and sanitize inputs to mitigate prompt injection patterns, repetition attacks, encoding tricks, and adversarial token sequences. Route detected stress patterns to controlled fallback flows.
  • Constrain generation behavior: Use structured output formats, temperature controls, max token limits, and deterministic decoding where appropriate to reduce variance and unpredictable outputs in critical workflows.
  • Perform targeted fine-tuning: For self-hosted models, retrain using datasets that include repetition traps and hallucination patterns to improve internal resilience. This is not applicable to provider-managed APIs.

​
Risk

Governance/Compliance Low robustness violates key expectations for safe and reliable AI operation:
  • OWASP LLM Top-10 (LLM05: Insecure Output Handling): Requires safeguards against unsafe or hallucinated outputs.
  • NIST AI RMF (MEAS-2, MAP-1): Expects validated reliability and resilience to abnormal or adversarial inputs.
  • EU AI Act (Art. 9): Requires risk controls that keep models stable under unexpected conditions.
Security Unstable models:
  • Hallucinate package names
  • Fail in divergence stress tests
  • Become inconsistent under repetitive stress
  • May emit fabricated or harmful advice
This undermines trustworthiness and reliability in operational systems.

​
Explanation

Robustness metrics capture stability under unusual or adversarial patterns such as divergence, hallucination, and abnormal sequences. A low score indicates fragility in reasoning and consistency.

​
Specifications

Trigger: evaluation.robustness_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