August 18, 2026 WitnessAI Release

WitnessAI now supports Microsoft Copilot Cowork, with expanded protocol compatibility to ensure reliable coverage across Cowork sessions.

WitnessAI now supports Microsoft Copilot Studio, including prompt and response observability for Copilot Studio agents.

WitnessAI now supports ChatGPT Work, providing observability and policy enforcement for organizations using this workspace.

WitnessAI now supports Databricks Genie, including prompt, response, and tool call observability.

WitnessAI now supports Devin, including prompt, response, and tool call observability.

WitnessAI now supports Antigravity, including prompt, response, and tool call observability.

WitnessAI now supports policy enforcement for AWS Bedrock, as well as prompt, response, and agentic observability for models accessed through Bedrock.

The Windows endpoint agent package for Microsoft Intune has been restructured to separately include platform scripts and remediation scripts, so organizations can deploy and monitor the agent using Intune's Remediations workflow in addition to standard platform script deployment.

Load performance for the Console dashboard and Discovered Apps page has been improved, particularly for organizations with large volumes of application and usage data.

Changing filters or date ranges on dashboard views now shows a loading indicator while data refreshes, rather than pausing the view on screen.

Improved resilience of alert processing during periods of elevated backend latency, reducing the likelihood of processing errors under heavy load.

Fixed an issue where a specific type of tool query could in some cases bypass blocking when an Agent Tool policy was configured in allowlist mode with a small number of allowed servers. Fixed an issue where incomplete tool metadata for certain agentic sessions could in some cases affect the accuracy of tool policy evaluation. Fixed an issue where attachment contents could in some cases fail to be redacted according to policy. Fixed an issue where guidance text describing supported attachment file types and size limits for Guardrails could be inaccurate.

Fixed an issue where certain values — including Social Security numbers, SWIFT codes, credit card numbers, and database credential formats — could occasionally be flagged as sensitive data when they were not genuine matches, reducing false-positive alerts across several Data Protection guardrails. Fixed an issue where certain workplace-related conversations could fail to trigger the intended topic classification alert.

Fixed an issue where adding sensitive data to a prompt could in some cases cause unrelated code block content in the same conversation to be masked unnecessarily. Fixed an issue where selecting an alert could in some cases fail to scroll to the associated prompt.

Fixed an issue where the agent count on the Discovery view could in some cases not match the detailed agent list for the same filter. Fixed an issue where the user count on the Discovered Apps page could in some cases be inaccurate. Fixed an issue where the list of applications for a user could in some cases fail to fully populate. Fixed an issue where web-fetch tool calls in Claude sessions could fail with certificate errors.

Some users may experience slower response times when WitnessAI inspects traffic from agentic AI applications.

Impact: Response delays may be noticeable in latency-sensitive agentic workflows.

Workaround: Contact Customer Support to review your policy configuration and receive recommendations for improvement.

Status: Under investigation. We will provide updates as optimizations are released.

Blocking an individual tool that has a broader, capability-redundant alternative may not fully prevent equivalent access if another tool can accomplish the same task. Coverage for these cases is planned in an upcoming release.

No breaking changes in this release.

Please contact Customer Success if you have questions or need support with this release.