Updates

​

Summary

​

New

• Added search functionality to dropdown menus with more than 10 items for faster navigation when configuring charts and managing resources.

​

Improved

• Added labels to the identifier and description fields for organization settings.
  • Added a hover state improved discoverability discover.
• Improved visual heirarchy between status pill components compared to general tags.

​

Fixed

• Fixed broken documentation links in Org Settings for Members and Workspaces tabs following documentation restructure.
• Fixed an issue where closing toast notifications would also close open modals.

​

Summary

​

New

​

Multi-Organization Support & Role-Based Access Control

​

What’s New

• Complete data isolation between organizations.
• Invitation-based user onboarding with secure, time-limited tokens.
• Three predefined roles: Owner, Contributor, and Reader.
• Organization Owners have full administrative control over users and settings.

• Share workspaces with individual users or your entire organization.
• Three permission levels: Owner, Contributor, and Reader.
• Workspace Owners have full control including deletion and ownership transfer.
• Contributors can create and edit projects and runs.
• Readers have view-only access for stakeholders and auditors.

​

Security by Design

• Complete data isolation between organizations with no cross-organization access.
• Invitation-based onboarding with cryptographically secure tokens that expire after 7 days.
• Principle of least privilege enforced at every level—Platform Admins manage organizations but cannot access workspace data.
• Private default workspaces for every user, accessible only by explicit sharing.

​

What This Means for You

• Collaborate on offensive AI projects with confidence that access is controlled.
• Share results with stakeholders using read-only permissions.
• Maintain separation between different client engagements or internal projects.

• Invite team members and assign appropriate roles.
• Create workspaces for specific projects or initiatives.
• Monitor who has access to sensitive red teaming work.

• Invite team members and assign appropriate roles.
• Create workspaces for specific projects or initiatives.
• Monitor who has access to sensitive red teaming work.

​

Learn More

• RBAC & Collaboration Introduction — Start here for a comprehensive introduction.
• Key Concepts — Deep dive into roles, permissions, and access control.
• Manage Your Organization — Guide for Organization Owners.
• Collaborate with Workspaces — Learn to share and manage workspace access.
• Permission Matrix — Quick reference for all role capabilities.

​

Improved

• Workspace names now follow the exact casing specified by the user.
• Enhanced the data exploration workflow with several improvements:
  • Users can now export all metrics without knowing exact metric names upfront (default behavior).
  • Users can still filter to specific metrics when needed for focused analysis.
  • Export APIs no longer error when runs have different sets of metrics (heterogeneous runs supported).

​

Fixed

• Removed the redundant chats tab, as chat rendering is now available in the task table and details pane.
• Fixed an issue where users might get redirected infinitely while trying to access Crucible.
• Fixed an issue that prevented customers from submitting feedback.

​

Summary

​

New

• Added support for published capability images with dependencies, enabling easier access to agents, evals, and attacks through the SDK.
• Introduced a new top navigation bar for platform-wide navigation, replacing the previous side navigation.

​

Improved

• Changed the summary display to a toggle control
• Made accordion headers consistent throughout
• Improved anchor link navigation for easier content access
• Improved metrics alignment for better readability
• Improved the chat tab appearance when no chats are present
• Improved the toggle, legend, and tags in the trace view

​

Fixed

• Fixed a minor bug that prevented part of the navigation menu from displaying.
• Fixed an issue that prevented large chat threads from displaying in Strikes.
• Fixed incorrect sorting in the table on the Strikes projects page
• Fixed an intermittent error in the table chart configuration
• Fixed scrolling functionality when hovering over charts

​

Summary

​

Improved

• Added clear status indicators for Strike runs, now displayed next to the run name on the run page.
• Added nicely rendered AI chat messages created from the new dn.Agent SDK methods.
• Expanded the width of anchor links in Strikes for easier navigation.
• Resolved issues with incorrectly formatted JSON on large outputs.
• Ensured chats from older Strikes runs remain accessible in the new UI.

​

Fixed

• Fixed an issue where tool information was not being displayed in the chat list.

​

Summary

​

New

• We’ve added AI Red Teaming tools to Strikes to systematically probe security and safety vulnerabilities in AI systems. These provide ready-made adversarial attack techniques and a flexible API for building your own attacks, targets, and analysis.

• Prebuilt Attacks: We introduce two prebuilt attacks that are both effective—requiring fewer queries to succeed—and efficient—achieving high jailbreaking success rates (measured as a percentage). The metrics used to evaluate these attacks include the total number of queries required and the overall success rate.
  • tap_attack - Tree of Attacks with Pruning, a structured tree based attack that expands candidate prompts and prunes off topic and low-scoring branches early, delivering high success rates while minimizing wasted queries and compute.
  • goat_attack - Graph of Attacks with Pruning, a graph traversal approach that carries the attack context throughout the nodes in a non-linear manner.
• Attacks Evaluation: Provide the ability to run multiple attacks with the same experimental settings and produce side-by-side comparisons across measurable metrics (e.g., total execution time, total trials made, attack success rate)
• Attack Budget Control: Ability to set a threshold or stop when max_trials is reached. The attack stops when either the scorer threshold or the max_trials limit is reached, allowing control over the attack within a finite budget.

• SimBA Attack: Simple black-box attack where no gradients are required that tries small perturbations in different directions and keeps the ones that lower the model’s confidence, quickly finding a subtle adversarial tweak with few queries.
• HopSkipJump Attack: A decision based, query efficient, black-box attack for evading ML models, particularly classification techniques.

• LLMTarget: point at any Rigging-compatible LLM (OpenAI, Anthropic, Groq, etc.).
• CustomTarget: wrap arbitrary functions or HTTP APIs (e.g., vision endpoints) with explicit control of the input parameter.

• Every run returns a StudyResult with best_trial, full trials, and failure/pruning details.
• Export rich telemetry via .to_dataframe() or .to_jsonl() for audit, trending, and post-mortems.
• All data ships to Strikes platform for durable storage of attack data, visualization, audit support, and advanced analysis.

​

Improved

• Improved the ability to analyze Rigging batch calls in Strikes.
• Added tool call descriptions directly to the tasks table in Strikes for easier reference.
• Enabled analysis of individual prompts when using Rigging with Strikes.
• Embedded chats directly within the Strikes tasks table for quicker access.
• Enhanced browsing of LLM chats in Strikes to make navigation easier.

​

Summary

​

New

• Added the ability to search the Tasks table. You can now search for data within your Strikes runs.

​

Improved

• Updated the UI to only return 10,000 runs for each project. Larger queries will require the API.
• Improved the scrolling behavior for charts in Strikes.
• Improved handling for long metric names for charts in Strikes.
• Improved the rendering for tools in Strikes.
• Improved the table sorting experience in Strikes.

​

Fixed

• Fixed a bug for enterprise customers that caused the latest release to not be available for download.
• Fixed a minor issue with searching tasks in Strikes.

​

Summary

​

New

• Added the ability to render data directly inside the tasks table, removing the need to open a popover or sidebar.
• Refactored several aspects of the tasks table in Strikes for better performance and usability.
• Centralized object detection logic into a utility and assigned enum values across all schemas.

​

Improved

• Users can now migrate ClickHouse aggregating tables during maintenance windows without engine compatibility conflicts. The offline migration tool correctly handles both MergeTree and AggregatingMergeTree engines, with proper environment-specific selection for local vs. shared cluster configurations.
• Reduced token refresh frequency:
  • Enhanced flexibility: Users can customize token duration to fit their needs.
  • Improved developer experience: Fewer interruptions during long-running operations.
  • No action required: Existing integrations automatically benefit from longer default durations, while users needing shorter ones can set a custom parameter.

​

Fixed

• Fixed wrapping for long strings of text embedded inside arrays on the tasks table by adding character-level breaking, improving readability.
• Fixed string formatting issues in the tasks table by preventing pre-formatting defaults. Customers using Python format strings will now see cleaner, more accurate rendering.
• Fixed an issue where pending runs failed to update in the UI until completion.
• Fixed an issue preventing multiple runs from being deleted at once, improving workflow efficiency for managing large experiment sets.

​

Summary

​

New

• Added a checkbox to select metrics not currently in a chart, and added auto-generated charts for those metrics. You can now add multiple charts with one click.
• If you’re an enterprise customer, you can now download a self-hosted platform installation from the Platform Downloads section of the Account page.

​

Improved

• Charts now display automatically when the project page loads.
• Improved the overall flow of Strikes data exploration:
  • The run page now shows only the run summary.
  • To see task-level details, there is now an Explore Tasks button above the summary.
  • The vertical split layout is now a horizontal split for better content flow.
  • Individual tasks now have their own URL, making it easier to share findings.
• Improved documentation for Strikes with self-hosted environments.
• Added hyperlinks to the Terms and Conditions and Privacy Policy in a new “Legal” section on the Account page to provide transparent access to important notices.
• Updated the Contact page with a clear distinction between the support email and the customer success email and clarified what their purposes are. You should have a better experience on smaller devices when viewing this page.
• Removed the requirement to include an aggregation option in Strikes charts.

​

Fixed

• Fixed an issue that caused self-hosted Platform deployments to fail to regenerate API keys.
• Fixed an issue that impacted in-browser chats with LLMs.
• Fixed an issue that prevented parameter columns from correctly toggling on the Strikes Runs page.

​

Summary

​

New

• Enabled media rendering. The Strikes UI can now render audio, video, and images logged as inputs or outputs using the Dreadnode SDK.

​

Improved

• Added missing event data from run schemas to improve charts for optional lineage graphing.
• Improvements to the scatter plot in Strikes including an ability to select an aggregate.
• Performance improvements for large projects.

​

Fixed

• Fixed content security policy issues, including enabling media logging and viewing (images, videos, and audio) and resolving potential header errors.

​

Summary

​

Improved

• Cleaned up our email communication process for new users.
• Improved copy and fixed errors in challenge names and descriptions.
• Increased difficulty on some challenges to align with model improvements.
• Improved responsiveness on the login screen.
• Updated challenge notebooks for standard alignment and consistency.
• Fixed a traces parsing bug between the SDK and platform.
• Improved default server handling during the configuration process.
• UI bug fixes and charting improvements across the product.

​

Summary

​

New

• Added metric KPIs to run details for better performance tracking.
• Added an empty landing page for Strikes to prepare for upcoming features.
• Implemented waitlist functionality for managing access to new features.

​

Improved

• Added artifact retrieval functionality to improve access to test results.
• Added links to Strikes examples on GitHub for easier reference.
• Replaced trace view score bubble with input, output, and metrics columns for clearer data visualization.
• Improved chart rendering and fixed overflow behaviors in tables.
• Enhanced rendering for flow graph nodes to improve visualization.
• Inputs and outputs are now grouped by task for better organization of results.
• Enhanced rendering for flow graph nodes to improve visualization.
• Increased OpenTelemetry retries to improve stability.

​

Fixed

• Fixed sorting and filtering for user-supplied parameters in data tables.
• Addressed UI problems in charts, tables, and data visualization components.
• Fixed image generation issues that were causing broken code.
• Updates and bug fixes for challenge stability.

​

Summary

​

Improved

• Improved our documentation and added CTF Agent and NET Reversing Agent use cases.
• Improved the overflow rendering for trace tree.
• Added the ability to bulk delete rows of runs.

​

Fixed

• Fixed an issue with the row checkbox.
• Fixed an issue with Strikes project runs row selection.
• Fixed an issue with bulk deleting runs.
• Fixed minor bugs across the product to improve usability and create a smoother user experience.

​

Summary

​

New

• Released the Voyager challenge.

​

Improved

• Refactored the challenge architecture.
• Updated our challenge environments and configurations.
• Improved the instructions for running a challenge.
• Improved styling components.

​

Fixed

• Fixed logger issues in the Popcorn2 and Librarian2 challenges.
• Corrected the challenge variable definitions in notebooks.
• Fixed a submit_flag function bug in Canadian Eh challenge notebook.
• Fixed the submit flag endpoints in the notebook to ensure proper functionality.
• Resolved errors in the Hotdog challenge notebook.
• Corrected the PhantomCheque challenge tags for accurate categorization.