Release notes: Nov 28, 2025 (v.9.82.3)

AI Guardrails

Easily test scanner packages

We've improved the Playground to help you test multiple scanners within a single package more easily. This release introduces the ability to filter the scanner view and test an entire scanner package against a custom dataset.

• Improved package view — Filter the scanners view in the Playground to focus on a single scanner package. Then select the Test toggle to enable all scanners for testing at once.
• Test scanner package with a dataset – Upload a dataset and test it against all the scanners in a selected package. The latest version of each scanner is selected by default, or you can select any published or unpublished version of that scanner.
• Edit scanner package names – You can now directly edit a scanner package name by selecting Rename package from the three-dots menu. The scanner name is editable and you can save by hitting Enter.

View top 5 connections by usage

We redesigned the connections usage graph on the dashboard to show the top 5 connections by usage, giving better insight into the top talking models. Click View all to see the complete list when there are more than 5 models to show.

AI Red Team

Connect and protect any agent

We're excited to release a new Agent Project specifically for connecting and protecting agentic systems and workflows. AI Guardrails can connect any agent that uses OpenAI-compatible APIs, via a generated API key and session ID. 

• Simple onboarding: Use the agent project wizard to name the project, generate an API key, and add users and model connections.

• Session ID: When implemented, a session ID allows parsing of the agent runs into logical, connected flows and better audit trails.

• Multi-model agents supported: Connect one or many models, including public and private models.

• Enable scanners: Protect the agent from unauthorized and dangerous behaviors with pre-built scanner packages and custom scanners you define. Set scanners to block for critical risks, or audit for anomalies and agent drift. 

• View audit trails: All agent and scanner activity is logged in prompt history so you can easily identify issues or opportunities for fine-tuning.

Create a better custom intent with AI assistant

We added a simple AI assistant to custom intents to help you craft a more effective attack objective. The assistant is instructed on our best practices (Best practices for writing custom intents), which include:

• Focusing on one topic per intent
• Being direct and imperative
• Being concise
• Using proper spelling and capitalization
• Avoiding leading and trailing punctuation

Access the AI assistant by navigating to Attack campaigns > Create Campaign > Agentic Warfare. The AI assistant button is enabled when you start typing your prompt or when you edit an existing prompt. You can accept the AI assistant’s suggestion or ignore it.

December attack pack: Adversarial metaphor

This month’s release introduces Adversarial Metaphor, a black-box jailbreaking technique that tricks a model into providing instructions for harmful actions (like building a weapon or planning a malicious act) by disguising the query as a benign, philosophical, or creative task. As usual, the new attack package contains over 10,000 single-turn prompts that use 14 distinct attack vectors, providing protection from existing and emerging attacks.

Change a report name

We added the ability to change the name of a Red Team report before or after completion, so it's easier to identify. Click the three-dots menu, select Edit report name, rename it, and hit Enter. Scheduled reports that haven't run yet are not renamed.

Other enhancements

• Logs: We added a user filter to the logs page, allowing you to easily search and narrow down log entries by a specific user.
• Custom roles: To make the customer roles matrix easier to navigate, we've added collapse/expand arrows to each permissions block.

Bug fixes

• The patch request to disable a scanner did not revert the scanner mode to 'block', leaving an inconsistent state upon deactivation. - Resolution: Fixed.

• The Select All option in the select component did not correctly respond to changes in the visible options caused by the Search filter, leading to an inconsistent selection state. - Resolution: Fixed.

• In Outcome Analysis, the functionality to select/deselect individual scanners or use the Clear All button was not working. - Resolution: Fixed.

• The edit/version history view was missing breadcrumbs, making navigation back to the previous screen more difficult for users. - Resolution: Fixed.

• The actions count column was missing from the reports table. - Resolution: Fixed.

• A table header shadow was visible in the Manage Connections view, specifically in the top left corner of the first item in the list. - Resolution: Fixed.

• The task request was executed only for Org Admins and failed to execute for users with organization-only permissions. - Resolution: Fixed.

• GenAI scanner descriptions were visible on the Projects page under custom scanners, but are missing from the scanners list under Custom scanners. - Resolution: Fixed.

• Incorrect token count was displayed in the API tokens tab on the project page. - Resolution: Fixed.

• The organization settings tab was not rendering correctly on the settings page. - Resolution: Fixed.

• The UI displayed a redundant toast notification for every previous Default model change when a new model was selected, leading to multiple, cluttering notifications. - Resolution: Fixed.

Known issues

In Custom Roles, clicking on the name of an individual permission can cause it to disappear from the list. Refreshing the browser tab will reset the view.