May 19, 2026 WitnessAI Update

WitnessAI Release: May 19, 2026

New Features

Agentic Observability — Microsoft 365

• Microsoft 365 Copilot iOS — Prompt Support: Observability is now supported for the Microsoft 365 Copilot iOS thick-client app. Prompts and responses are captured for monitoring and policy enforcement.
• Microsoft 365 Copilot iOS — Attachment Support: Attachments sent through the Microsoft 365 Copilot iOS app are now captured and analyzed.
• Microsoft 365 Office Apps iOS: Observability support added for Microsoft Word, Excel, and PowerPoint on iOS. Prompts and responses from each app are captured separately and attributed per application.

Witness Anywhere — MDM Integrations

• Added Kace (macOS), Intune (macOS), JumpCloud (Windows), and JumpCloud (macOS) as script download options under Proxy Configuration, enabling streamlined deployment through additional MDM platforms.
• Added Workspace One (macOS) and Workspace One (Windows) as script download options under Proxy Configuration.
• JumpCloud integration support added for Witness Anywhere on Windows devices.

Feature Updates

Agentic Policy & Control

• Audit Trail — Enforcement Observability: The console now surfaces which MCP tools and servers were allowed or blocked for each prompt, providing a per-prompt enforcement audit view. Backend tracking records enforcement decisions per user per prompt to support audit trail requirements.
• Global Block Message: Administrators can now configure a global default block message that is delivered to end users when a policy blocks an action, consistent across the proxy and backend.
• Custom Allow/Block Messages for MCP Policies: Per-policy custom allow and block messages can now be stored and used during enforcement, enabling tailored user-facing messaging for different policy contexts.
• MCP Server Add — Single-Step Flow: Adding an MCP server to the policy list now creates only the server-level entry. Tools belonging to that server are no longer automatically injected into the tool list, giving administrators explicit control over tool-level policy.
• Policy Source Filtering: The Policies and Agentic Policy pages now surface the originating source (such as VSCode) for each policy. Policies can be viewed and filtered by source, destination, and guardrail with typeahead search support.
• Blocked MCP Servers/Tools in Conversations: The Conversations view now includes a panel showing which MCP servers and tools were blocked by policy for a given conversation.

Custom Data Types

• Administrators can now view and delete custom data types directly from the console UI, without needing to use the API.

Advanced Alert Filtering

• Conversations and alerts now support advanced filtering. Backend query endpoints and console filters have been updated to support search by additional attributes, enabling more precise investigation of prompts and alerts.

Agentic Observability

• Writer.com — Agentic Tool Call Observability: Agentic tool call activity from Writer.com is now captured and visible in the console.
• VSCode — Claude Opus Model Visibility: When a user switches models mid-conversation in VSCode, Claude Opus is now correctly surfaced in the Conversation Agentic Context View.
• VSCode Codex Models — WebSocket Support: Observability support added for VSCode Codex models, which now use WebSocket-based communication.
• Cursor — LLM Response Timing: First-byte and round-trip response timing metrics from the LLM provider are now captured and displayed for Cursor conversations, consistent with other supported applications.

Conversations & Analytics

• Request and Response Payload Size: The console now displays request and response payload sizes for each conversation, helping correlate latency with data volume.
• Policy Creator Tracking: Policies now record the user who created them. The policy list supports filtering by creator, making it easier to manage policies in multi-user environments.

Supported Applications

• Claude Cowork and Claude Code Desktop: Both applications are now recognized as supported apps in the WitnessAI console.
• Alert Details: Conversation details are now rendered in the alert popover, giving administrators richer context when reviewing alerts without leaving the current view.

Performance, Reliability, & API Updates

• Kafka Resiliency: Improvements to Kafka messaging resiliency and throughput, reducing risk of data loss under high-load conditions.
• LLM Timing Header Accuracy: Fixed missing LLM first-byte and round-trip timing values on empty-body responses, ensuring accurate latency data is recorded for all response types.
• vLLM Support in Proxy: vLLM traffic is now routed through the parserlib inspection pipeline, enabling policy enforcement and observability for vLLM-served models.
• Access Control Hardening: Resolved a broken access control issue in the private applications listing endpoint where some lower-privileged users could access data beyond their authorization level via direct API calls.
• Witness Anywhere — Device Registration Logging: Failed device registration attempts are now recorded in the database, enabling administrators to audit and diagnose registration issues.
• User Profile Completeness: Some optional user profile fields were not being persisted at account creation time. This has been corrected to ensure complete user records.
• Settings — Roles Permissions Table: The Roles permissions matrix in Settings has been migrated to an updated table component, improving rendering performance and consistency with other console tables.

Fixed Issues

Data Protection

• Fixed an issue where data protection guardrails would only trigger for the first custom data type in a policy; subsequent custom data types are now evaluated.
• Fixed an issue where alerts for custom data types were always generated at High severity, regardless of the alert level configured in the policy.
• Fixed an issue where changing a data protection policy from Block to Warn caused prompt submission to stop working in Claude desktop and web.
• Fixed an issue where updating a data protection policy in Google Sheets caused both Block and Warn messages to appear simultaneously in the Gemini Sidebar.

Proxy & Observability

• Fixed an issue where some Microsoft Copilot Enterprise "research" prompts were intermittently not being captured in Witness due to a query ID change.
• Fixed silent message drops during WebSocket rotation in the VSCode Codex integration.

Policies

• Fixed an issue where the MIP policy purpose field was not being saved.
• Fixed group-based RBAC controls for the prompts-guardrails API that were failing for some role configurations.
• Fixed an issue where a blocklist configuration was unexpectedly removed.

Console & Display

• Fixed an issue where blocked prompts were not visible in the console when the "enhanced context for agentic" setting was enabled.
• Fixed a user count mismatch between the MCP server list and the detail flyout in the Discover view.
• Fixed tool execution count displaying as 0 in some conversations even when tools were invoked.
• Fixed an issue where duplicate prompts that triggered guardrails did not display the guardrail message.
• Fixed sidebar navigation flickering during page transitions in the console.
• Fixed an issue where sidebar navigation became unresponsive after the Analytics dashboard fully loaded.
• Fixed missing loading indicator when filters were changed on the console homepage.
• Fixed Anthropic (Claude) missing from the Available Providers list.
• Fixed an issue where some conversations were failing in QA environments due to a console version incompatibility.
• Fixed an inability to delete a policy from the console UI when using the typed confirmation field.
• Fixed out-of-memory errors caused by some large attachment processing operations.
• Fixed prompt text not displaying correctly in the Conversations view when the prompt included a file attachment.

Known Issues

Breaking Changes