This is a SaaS-only release.
Inference Red-Team
June attack pack & new attack vector
This release includes our June attack pack, with over 10,500 new malicious prompts, and a new attack vector: style injection.
Style injection is a jailbreak technique that works by adding specific writing or formatting rules (e.g., "do not use punctuation" or "use no more than 5 words per sentence") to the prompt, as a way to distract the model from using its standard refusal language or phrase, and instead respond with an unsafe response that would ordinarily be blocked.
Sharp-eyed users will notice that we’ve changed the name of our monthly prompt packs to attack packs. This makes our terminology more consistent with current and future attacks.
Improved usability in the raw data table
We’re continuing to improve usability in the raw data table for Red-Team reports based on customer feedback. In this release you’ll see:
-
Better filtering and sorting capabilities.
-
Visual indication of the number of results in the table, before and after filtering.
-
Automatic loading of more results as you scroll down.
Keep the suggestions coming!
Campaigns are more visible
We’ve added a column for campaigns on the Reports page, and added a campaign link to both the UI report and the downloadable report. The campaign name link takes you to that exact campaign. The raw data view now shows columns for the report name and the campaign ID number.
Error messaging for reports
We’re in the process of improving Report error messaging so it’s easier to troubleshoot and recover from issues. In this release we’ve added a category for schedule errors.
A schedule error occurs when there is a problem with one of the target model or application connections during a scheduled campaign run. Users seeing this error should test the connection and fix any issues. When the connection is fixed, Red-Team will resume with the next scheduled run.
CalypsoAI Inference Defend
Custom scanner details
We’re building more functionality into custom scanners and, as a first step, we’ve added a detail view that shows the tags and full scanner description. Click on the scanner name in the table to view its details.
Bug Fixes
-
Customers in the US may have noticed that the Signature attack pack dates went from December 2024 to April 2025. This was incorrect and due to a timezone parsing issue. We’ve made a fix and all users should see the correct dates, starting in January 2025. The attack packs are exactly the same – only the months have changed.
-
We’ve introduced a limit of 100 characters to prevent very long campaign names from causing layout issues on the Campaigns table.
-
We fixed an issue in the Chat UI that caused new messages to be missing when users returned to a previous chat via a template.
-
We’ve corrected an issue that arose when a project user removed a scanner package in which some scanners were enforced at the global level. Previously, it would appear that the package was successfully removed, which was not the case since enforced scanners settings cannot be changed. In this release, the project user will see instead get a message explaining why the package cannot be removed. Project users can still disable any non-enforced scanners in the package if they need to.
-
On the Dashboard, we’ve moved the Filter button to the left to be consistent with the rest of the UI.
Known Issues
We’re aware of an issue when creating a Red-Team attack campaign – when users click on an existing campaign, then create a new campaign, the latest attack vectors may be unselected. If this occurs, click the All attacks checkbox to select everything.