Jan 29, 2026 - SaaS - v9.133.3.3

AI Guardrails

UI enhancements

Projects

• Users can now filter projects by type: agentic, app, bot, chat.

• We added a search bar to the Project > Add Scanners page, making it easy to quickly find a specific scanner or scanner package.

Scanners

• Scanners with no published versions appear disabled, to better distinguish them from active scanners. Click on the scanner to View versions and select a version to publish.

• We added icons to the Scanners page to better distinguish packages from individual scanners.

• The Rename package option is now Edit package. Users can not only rename the package in this view, but can also add and remove scanners.

• Unnecessary filter and sort controls have been removed from the CalypsoAI scanners list and the search field and Build button are aligned to make the UI more compact.

• The Custom scanners list has a new filter dropdown with more options and controls.

• We improved bulk selection of custom scanners and packages by making the checkboxes visible at all times.

Playground

• When testing a scanner package with a custom dataset, users can now see the aggregated F1 score for the whole package, not just the individual scanners inside the package.

AI Red Team

February attack pack: Invasive context engineering

This month's update introduces Invasive Context Engineering, a sophisticated attack vector designed to test how LLMs stay aligned to their intent during long conversations.

This attack exploits a weakness in LLM alignment—the longer the context, the weaker the influence of system instructions and directives. Injecting small, repeated instructions into a request can anchor the model’s attention on those instructions instead of its initial programming, making it vulnerable to adversarial control or alignment drift.

The February pack contains over 10,000+ new attacks, comprising 1,500 plaintext prompts across 58 intents and 7 converters. This update adds 123 invasive context engineering prompts to our arsenal.

Other Red Team enhancements

• Attack vector names in the mitigation actions list now use sentence case, aligning them with the campaign panel and our styling guidelines for improved consistency and readability.

• The Reports table was getting cluttered, so we removed the campaign and schedule columns and made the progress bar more compact.

• Agentic fingerprints is no longer in beta and is ready for general use.

Bug fixes

• Users trying to set up or connect to Anthropic models may have encountered an API error due to an API change at the provider. Resolution: Fixed.

• After unpacking a remediated scanner package, the test link on the Remediation page led to an “Unexpected error” page due to the scanner package not being found. Resolution: Fixed.

• The org_id taken from context in the bot service did not match the chatbot’s org_id, causing Slack bots to be broken for any organisation outside the system/admin org. Resolution: Fixed.

• Project scanner counts became inconsistent between the UI, API, and scan results under rapid scanner updates, causing non-zero counts to persist after scanners were removed. Resolution: Fixed.

• Multiple active entitlements could be created for a single organisation due to a race condition, resulting in a broken organisation state. Resolution: Fixed.

• Attack campaign names exceeding 100 characters ran off the screen and could not be fully viewed due to missing text wrapping or horizontal scrolling. Resolution: Fixed.

• Clicking any option within custom scanner filters caused the playground panel to close unexpectedly. Resolution: Fixed.

• The Settings save button became stuck in a loading state when the retention confirmation dialog was dismissed by clicking the close icon or outside the popup. Resolution: Fixed.

• An unclear “Unexpected error” message was shown when rescheduling a report with a past start time, instead of indicating that the start time must be in the future. Resolution: Fixed.

• The data retention input allowed excessively large values to be entered, leading to invalid numbers (e.g., scientific notation) and backend errors due to integer limits. Resolution: Fixed.

• Regex test highlighting did not apply on the first valid match, only appearing after additional input was entered and then persisting when the extra text was removed. Resolution: Fixed.

• In the Playground, a “Text Copied” message appeared prematurely when hovering over a prompt, before the user clicked the copy icon. Resolution: Fixed.

• The Chat page unexpectedly formatted pasted text containing bullet points as code snippets, despite no markdown syntax being present. Resolution: Fixed.

• A 500 error occurred when attempting to enable a scanner that had been previously unpackaged from a scanner package, preventing the scanner from being enabled individually. Resolution: Fixed.

• The Edit Package panel incorrectly displayed a “Search” icon in the package name input field, causing confusion since the field is used to edit the package name, not perform a search. Resolution: Fixed.

• Blocked response warning messages in Chat did not include the associated promptId, making it harder to identify the affected prompt. Resolution: Fixed.

• Tables in chat responses were not rendered correctly, displaying excessive blank lines instead of properly formatted tables. Resolution: Fixed.

• Redaction was not applied to response content when using OpenAI compatibility, causing sensitive data to remain not redacted in the returned payload despite being redacted in logs. Resolution: Fixed.

• A “global” project briefly appeared on the Projects page after clearing filters. Resolution: Fixed.

• Downloading dataset results failed with an unexpected “file not found” error, preventing users from accessing completed dataset outputs. Resolution: Fixed.

• On the Audit Logs page, the event name was incorrectly displayed beneath the user instead of in its designated field. Resolution: Fixed.

• The Chat side panel did not use available vertical space after the footer was removed, resulting in a smaller-than-necessary chat area. Resolution: Fixed.

• Selecting a user in the filters panel caused the Datasets panel to close unexpectedly. Resolution: Fixed.

• Error messages in the Connections form could be cut off when too long, preventing users from viewing the full message. Resolution: Fixed.

• Scanner names were unnecessarily truncated despite sufficient space to display the full name. Resolution: Fixed.

• The Reset action in Projects > Scanners > Edit Scanner Response behaved unexpectedly by closing the text box instead of restoring the original text, and the reset change was not immediately reflected due to a caching issue until the panel was reopened. Resolution: Fixed.

Known Issues

• When filtering custom scanners by type, scanners in packages are not included in the results.

• In the Attack campaigns table, reports that are scheduled for the future are incorrectly showing a Last run time.

• In the Playground, datasets with the wrong case in the column head (Label instead of label) are allowed to run, but the job will fail.