Version 2.3.2 | June 09, 2025
Overview
ZBrain v2.3.2 delivers targeted enhancements and new feature additions to the following areas:
Application security: During app creation, configurable Guardrails enforce prompt-level safety aligned with compliance requirements.
Flow access control: Role-based access ensures secure, flexible use of flows.
Evaluation monitoring: Track evaluations with query-level logs and detailed permission controls.
Together, these updates strengthen ZBrain’s foundation for building secure, scalable, and governed AI solutions across complex enterprise environments.
New features
Apps
Guardrail configuration for secure app interactions
ZBrain now empowers users to implement guardrail validations, including input and jailbreak checks, directly during app creation or editing. These checks help ensure that user prompts are screened for safety and policy compliance before they reach the LLM, maintaining secure and controlled AI interactions.
Key capabilities:
Guardrail types:
Input validation: Detects and blocks unsafe, offensive, or non-compliant prompts using predefined safety rules.
Jailbreak detection: Prevents prompt injection or prompt manipulation attempts designed to bypass restrictions.
Configurable at the app level:
Users can selectively enable one or both options using a simple toggle interface during app configuration.
Guardrail settings are saved with the app and can be edited at any time.
Dynamic YAML support:
If custom guardrails are configured, ZBrain automatically generates and stores a dynamic YAML file.
Chat API integration:
Chat requests reference the YAML via a
YMLURLparameter referencing guardrail config stored in AWS S3 ({app_id}/config.yml).Guardrail settings are fetched from the content microservice at runtime if applicable.
Prompt evaluation flow:
If guardrails are disabled, prompts are sent directly to the LLM.
If guardrails are enabled:
Prompts are first validated.
If there are no violations, the prompt proceeds to the LLM.
If violations are detected:
Input checking → blocks known unsafe prompts.
Jailbreak → blocks exploitative prompt patterns.
Robust error handling:
If the YAML file is invalid, unreachable, or malformed, the system returns a clear and descriptive error message to the user or API consumer.
Improvements
Flows
Flow triggering via personal API keys
ZBrain now supports role-based flow execution via personal API keys.
Key enhancements:
Users assigned the Builder or Operator role by an Admin for specific flows can now trigger those flows using their own API key through the associated webhook URL.
Ensures secure and permissioned flow execution aligned with enterprise governance.
Gmail integration as a custom trigger
ZBrain Flows now includes Gmail integration as a native custom trigger, enabling users to:
Connect Gmail accounts
Monitor incoming messages
Initiate downstream actions based on email activity
Evaluation framework
Query-level monitoring and structured input capture
ZBrain now lets users turn on monitoring for individual queries within a session for both the app and the agent. The result is detailed, end-to-end traceability for each AI interaction on the platform.
Key enhancements:
Multiple queries monitored per session:
Each query can be independently monitored without losing session continuity.
The ‘Monitor’ button remains active for all queries unless they are already being monitored.
Users cannot create a new event for a query that is already under monitoring.
Detailed logging:
Each monitored query is linked to its Chat ID and independently logged, preserving token counts and evaluation metrics.
Structured input support:
Monitoring now handles inputs in structured formats, including plain text, uploaded files and file references (bucket, path, file type metadata).
Cross-module support:
Monitoring enhancements are extended across apps, agents and reasoning modules.
Secure access controls:
Monitoring permissions are enforced through user management API integration, with middleware applying role-based CRUD controls to every resource.
Last updated