August 14, 2026

Risk assessments now include a heatmap that plots every vulnerability against every attack method, colored by fail rate, with row headers that stay pinned while you scroll. Click any cell to drill straight into the test cases behind it, so "which attack breaks which defense" stops being a spreadsheet exercise. Your weakest cell is now the brightest one. Heat seeking, fully guided.

A new tab on risk assessments plots turns against attacks as a step-down survival curve, showing exactly how long your model keeps refusing before a multi-turn attack wears it down. Test cases also surface why an attack stopped early, so you know whether the model held the line or the attacker ran out of turns. Everyone breaks eventually—now you know on which turn.

Custom dashboards now speak the test run data model, and widgets can segment by metric—pass rates, median scores, and per-metric breakdowns land in tables right next to your traces and spans. One dashboard for offline evals and production, no more tab juggling. Run the numbers, literally.

The test run page now supports saved views, so the filter combination you rebuild every morning can be saved once and reopened forever. Share the setup, skip the setup. A room with a view, permanently booked.

Governance policies can now inherit from base policies, with controls from the base applying live to every child. Define the org-wide rulebook once, extend it per team, and stop copy-pasting controls between policies. Inherit the win.

Control filters now show which resources actually exist across your organization and how many projects carry each one, so scoping a control stops being a guessing game. Populated dropdowns, real counts, zero archaeology. Everything under control.

The Bedrock integration now supports IAM role-based access: spin up the role from our CloudFormation template, and we assume it with a per-organization external ID—no long-lived AWS keys pasted anywhere. Your security team can finally stop side-eyeing the credentials page. Role model behavior.

You can now choose the model that powers conversation simulation and red teaming attack simulation, at the organization or project level. Run your simulations on the provider you trust (or the one your provider policy allows). Sim-ple as that.

Press ⌘K anywhere on the platform to pull up a search bar and jump straight to what you need—no sidebar spelunking required. The fastest route between two pages is now two keystrokes. Special K, zero sugar.

Every platform page now ships with a guided tutorial, so new teammates learn features where the features live instead of in a docs tab they'll never reopen. Onboarding that walks, so you can run. Take the guided tour.

Metrics can now be composed as DAGs: chain steps together in a visual builder, with validation, versioning, and full API support for pushing and pulling definitions. Complex, multi-step evaluation logic without duct-taping metrics together in code. DAG-nabbit, it's good.

Self-hosted deployments now ship with the MCP server, so on-prem customers get the same coding-agent superpowers as the cloud. Your agents, your VPC, no exceptions. On-prem, on point.

Connecting your coding agents over MCP now runs through a proper OAuth consent flow with project discovery built in—authorize once, pick your project, no API keys pasted into config files. Consent screens: boring, correct, finally here. Auth-orized personnel only.

AI Connection streaming events can now be matched by payload type, so SSE streams with mixed event shapes parse cleanly instead of hoping every frame looks the same. Point at the type path, get the right field every time. Strong type energy.

Vertex AI now supports global and multi-region endpoint locations, and Anthropic models on Vertex route through the right publisher with catalog-accurate pricing. Claude on Vertex, from anywhere on the map. Think global, act model.

Test run graph card tooltips now render above everything instead of getting clipped by the card next door. Hover, read, move on—no more tooltip peekaboo. Stay on top of it.