October 2026 release notes for Test Cloud.

Prompt inputs and LLM-generated outputs are no longer saved in the audit logs by default for new AI Trust Layer policies. Input/output saving is disabled when you create a policy, so LLM calls governed by the policy appear in the audit logs with their inputs and outputs shown as Blocked by Policy until you enable it. Existing policies keep their current setting.

For details, refer to Viewing audit logs and Settings for AI Trust Layer policies.

Claude Sonnet 4 (anthropic.claude-sonnet-4-20250514-v1:0) is no longer available on the UiPath platform. UiPath removed it on 2026-10-13, following its retirement by Amazon Bedrock. Claude Sonnet 4.6 (anthropic.claude-sonnet-4-6) is the replacement.

UiPath-managed models are migrated to Claude Sonnet 4.6 automatically, with no action required on your side. Configurations that bring your own models through AI Trust Layer and reference Claude Sonnet 4 are invalidated. If you have not updated them to Claude Sonnet 4.6, they fall back to UiPath-managed models.

For current model availability across the UiPath platform, see AI features and model routing. For more information, see the LLM model deprecation timeline.

Model Hub now shows the status of Bring Your Own (BYO) configurations, lets you check the status of enabled configurations, and lets you enable disabled configurations. You can also optionally display the Folder column and see more context when deleting a BYO configuration. For details, see Managing models in Model hub and Configuring LLMs.

Private (Relay) connections for the Snowflake connector and for SQL Server in the Database Hub connector now run through the Relay on-prem executor. The JDBC drivers ship inside the executor, so no connector libraries or dependencies directory are needed. Requires Relay client 26.4.4 or later. Other Database Hub database types remain unsupported over Relay.

For details, see Supported TCP-based connections.

Relay client 26.4.4 adds the following:

For details, see Command reference.

Organization Administrators can configure a governance policy in AI Trust Layer that lets users automatically acquire AI and Agentic usage top-ups instead of requesting them manually each time. This feature is being rolled out gradually and might not yet be available for your organization.

For details, refer to Settings for AI Trust Layer policies, Granting AI and Agentic usage top-ups, and Requesting AI and Agentic usage top-ups. For the Automation Ops release note, refer to the October 2026 Automation Ops release notes.

Allowlisting the Cloudflare 104.16.0.0/13 outbound IP range might be required for any UiPath service, not only Document Understanding as announced in the July 2026 release notes. If your firewall restricts outbound traffic and your network equipment does not support DNS-based allowlisting, allow this range to prevent traffic from being blocked. DNS-based allowlisting is recommended because Cloudflare IP ranges are broad and may change.

For details, see Configuring the firewall for Test Cloud and Legacy - Configuring the firewall for Test Cloud.

gpt-4o-2024-05-13 is no longer available on the UiPath platform. UiPath removed it on 2026-09-30, following its retirement by Azure OpenAI. gpt-5.1-2025-11-13 is the replacement.

UiPath-managed models are migrated to gpt-5.1-2025-11-13 automatically, with no action required on your side. Configurations that bring your own models through AI Trust Layer and reference gpt-4o-2024-05-13 are invalidated. If you have not updated them to gpt-5.1-2025-11-13, they fall back to UiPath-managed models.

For current model availability across the UiPath platform, see AI features and model routing. For more information, see the LLM model deprecation timeline.

You can now bring your own guardrails into the AI Trust Layer, using your organization's own safety vendor subscriptions. Agent guardrails are then evaluated by your provider, under your own region and vendor agreement. During preview, you can bring subscriptions to Azure AI Language, Azure Content Safety, or Noma.

This helps organizations that cannot send prompt content to a provider they have not vetted, that have data residency requirements to meet, or that already pay for a safety service and want to reuse it across their agents. Your provider's guardrails sit alongside the UiPath-managed ones and work the same way, with the same scopes and actions: agent developers apply them to individual agents, and administrators enforce them organization-wide.

Every evaluation is recorded in the agent trace together with the provider that ran it, so you can account for which checks fired and where. Each guardrail can also fall back to its UiPath-managed equivalent if your provider becomes unreachable, so enforcement never depends on a single vendor staying available.

For details, refer to Configuring guardrails.