Skip to content
  • There are no suggestions because the search field is empty.

Release notes: June 20, 2025 (v8.216.0)

New in this release: Red-team connection rate limiting, performance enhancements, table filters UX

Inference Red-Team

Rate limiting for connections

We’ve added the ability to rate limit individual connections when they are used for red-teaming. Some applications and models are not able to process attacks as quickly as CalypsoAI can send them, resulting in errors and unsuccessful attacks. By applying a rate limit, admin users can set a maximum number of requests per second to ensure attacks are getting through.

To access rate limiting, go to Settings and click on the provider and connection you want to rate limit. Then click Advanced configuration and enter an integer for the requests per second.

NOTE: Rate limiting only applies to Red-Team attacks. If the same connection is used in a Defend project, it will not be rate limited.

 

rate-limiting

Intent category description update

We added some explanatory text to the intent categories chart on the Red-Team report to make it clearer what the data represent.

intent-category-description

Inference Defend

Performance enhancements

  • We’ve changed the scans API to be more performant for customers using the verbose parameter. In earlier releases using this parameter at scale caused noticeable latency, which has been effectively eliminated in this release.

  • We’ve also improved dashboard loading times.

CalypsoAI Platform

Filters on tables

We’ve improved table usability in the product by unifying all filters into a button and dropdown menu. A small number badge indicates when filters have already been applied. You can find the new filters on:

  • Dashboard

  • Logs

  • Projects

filters

Bug Fixes

  • Users reported a problem downloading Red-Team reports with a large volume of attacks. This issue has been identified and fixed.

  • We fixed a bug relating to the CASI score that caused models with no discovered vulnerabilities to receive a score less than 100.

  • We fixed an issue in the Campaigns UI that prevented users from clicking into the different attack techniques and converters.

  • Users trying to log into a disabled organization were not shown the correct error screen. This has been fixed.

  • Firefox browser users sometimes encountered an error when refreshing the page. The source of this issue has been identified and fixed.

  • We fixed a minor formatting issue in the Custom roles UI that occurred at narrower browser widths.

  • We fixed a Connections issue that caused an error to occur when the user unselected the option to “Enable access to model”.

  • In previous versions of the UI, it was possible to save a custom regex scanner with hidden spaces or returns that caused the scanner to fail. We’ve fixed the regex field to automatically remove spaces and returns at the end of the regex code on save.

  • In the Prompt history table, the download button is now disabled if the applied filters have zero results, preventing users from downloading an empty CSV file.

  • When running a campaign that mixed Agentic warfare and other attacks, the whole campaign could error out if just one agentic attack had an error. This has been fixed.

  • There was an issue that caused a page load error when selecting a campaign then refreshing the page, which has been fixed.

  • The Red-Team report PDF was only showing a partial page. This has been fixed.

  • Some pages that didn’t need it were showing a horizontal scroll bar. This was removed.

Known Issues

  • We’re aware of a bug customers encountered enabling the purge prompt history functionality and are working to resolve the issue for the next release.

  • If the API is used to set a package's scanners to both active and enforced at the global level, a project member can select/unselect the active toggle for that scanner in the UI, making it appear that they can override enforcement. This is a bug at the presentation-level only; enforcement is still applied in this case.