Skip to main content

3.23 Release Notes

3.23.1​

Improvements​

  • Improved, more efficient base model for output guardrails
  • Enhanced data generation module to reduce false positives
  • Streamlined policy deletion with better cleanup processes
  • Increased scalability of training datasets

3.23.0​

New Features​

  • [Alpha] Support for Japanese-language System Policy Compliance evaluations
  • [Alpha] Support for Anthropic Remote Cloud models
  • License-based access control for DynamoEval tests
  • Ability to cancel in-progress DynamoEval evaluations
  • Added Best-of-N as an adaptive jailbreak technique
  • Added new Projects feature for centralized guardrail and AI system management
  • [Alpha] Manage GPU resources by scaling policies up and down
  • [Alpha] Updated DynamoGuard PII Detection model with support for new entities - Account Number, Routing Number, Driver's License
  • [Alpha] Improved DynamoGuard Policy Management with support for more granular policy statuses, incremental data generation progress updates, and restart workflows for training and data generation failures

Improvements​

  • Improved System Policy Compliance test creation
  • Improved DynamoEval test error messages
  • Improved Default Prohibit Financial Advice Input policy
  • Updated policy data generation model with enhanced speed and data quality
  • Improved rate limit handling for policy data generation model
  • Restructured DynamoGuard Hallucination Policies into Input Relevance, Response Faithfulness, and Response Relevance