2025-09-17 – 2025-09-23 [Latest]

New Features

1. Risk Analytics suite has been added to Customer Intelligence:
   1. We are now monitoring and reporting on following security events:
      1. Fidelity - is the agent being true to its persona or not - a lot of failures here can indicate that the persona needs improvements.
      2. Jailbreak - tracking of attempted and successful jailbreaks.
      3. Prompt leakage - tracking of attempted and successful system prompt leaks.
      4. Profanity - tracking of nsfw user messages blocked by profanity filters.
   2. New features in the app:
      1. Risk Monitoring section on Analyze Overview tab.
      2. Four new Risk views on Customer Intelligence tab: Fidelity, Jailbreak, Prompt leakage, Profanity.
      3. Filters for all of the above.
2. Advanced Role-Based Access Control is now available to all users who have access to the CustomGPT for Teams suite:
   1. It is now possible to create a custom user role that applies just to selected agent(s).
   2. This role can have any combination of existing permissions, except account-level permissions, or be limited to chat-only access.
   3. API-key granulation is supported even for users with limited roles, and user can delegate only agents they have access to their own keys
   4. UI has been refreshed for setting up and assigning custom roles, as well as managing team members.
3. The new layout has been released for the Agents page, with the updated all of the list view, grid view and chat-only user role. Quick buttons are now dynamically chosen based on the agent's state, and take the user where they need to go: Build, Deploy or Analyze.

Improvements

1. It is now possible to instruct the agent whether or not it should ask for end-user confirmation before using a custom MCP action.
2. It is now possible to completely turn off the avatar displayed when there's no active conversation.
3. When using Agent Initiative for Live Chat deployment, it is now possible to choose whether or not to automatically start the conversation when a user clicks on the widget, or simply open the chat window.
4. When using Agent Initiative for Live Chat deployment, it is now possible to hide the starter question used for the initiative prompt from the starter questions list inside the chat window.
5. Auto-generated personas now include an instruction to insert a call to action or a hook for the end-user in every response.
6. The layout has been improved for the Demo agent welcome screen.
7. A modal with the best practices and limitations of Numeric Search feature has been added to the app.
8. “Sitemap” has been renamed to “Website” on the Sources page.

Bug Fixes

1. Fixed the issue where agents were sometimes returning “I don’t know” responses for queries like “Hello”.
2. Fixed the issue where the context-rich starter question was never displayed inside the chat window if there wasn’t at least one other starter question.
3. Fixed the issue where the context-rich starter question had slightly different styling from the rest of starter questions.
4. Add Custom Action dialogue now correctly displays errors for incorrectly filled input fields, and prevents users from proceeding in such cases.
5. Manage API keys modal will now not automatically close upon deleting an API key.
6. Fixed the issue with long website URLs breaking the UI on Sources page.
7. Fixed the issue where the agent would sometimes mention the custom context as one of numbered references.
8. Fixed the issue with numbered references sometimes being provided in the wrong format.
9. Numbered references in agents’ responses have been disabled for the GPT-5 model due to sub-par performance.
10. Fixed the CORS error that prevented some users logged in via IdP to log out of the system.