3.22 Release Notes

3.22.0

[Beta] Filter monitoring logs by action, end user id, and guardrail type
[Beta] Relabel the model response judgement in your System Policy Compliance evaluations and recalculate your evaluation metrics

Improved input guardrail training process, resulting in higher performance on benign prompts
Data generation stability and error handling improvements
Reduced storage overhead for default Output guardrails leveraging Low-Rank Adaptation (LoRA)
Improved output guardrail latency leveraging asynchronous inference
Improved safety harm categories for Adaptive and Static Jailbreaking evaluations
Improved judgements for Adaptive and Static Jailbreaking evaluations
Improved data generation for System Policy Compliance evaluations