The agent is done. The case isn’t.

Wiz’s Blue Agent investigates a threat and reaches a verdict. The analyst who acts on it works in Google Security Operations.

Google has owned Wiz since March 2026, and the integration is now Google’s own, but they are still two products. The verdict crosses through a playbook action that fetches whatever analysis exists at that moment. The integration’s sync job can carry status and comments both ways, and severity one way; the analysis itself never syncs. It lands in a ready-made widget, where an admin has added one, or in the step’s raw result.

The ready-made widget

Google’s content hub ships a widget for this action. Here it is, running Google’s own code on Google’s own example result:

It’s Google’s generic template: every field, in the order the data arrives. The verdict is plain text, no heavier than the ID. The time is 2026-06-09T16:12:58.253716Z, with no age. Confidence gets the same dots as severity. The conclusion is cut to one line. Nothing says what the agent used, or whether the case has moved on.

The raw result

Without a widget, the analyst opens the step’s raw result. In Wiz’s own launch screenshot, the analysis is dated 27 July and was fetched on 14 September: 49 days old, and nothing said so. A demo reusing an old threat is normal. The problem is what the screen does with it.

Here is that screen, redrawn, with five things to check on any screen that shows an AI’s answer.

1. What stands out. The most prominent time next to the answer is 14 Sep 19:00:34, in the dialog’s header and again in the side panel: when the step fetched the analysis. It has no time zone.
2. What format it’s in. The time the agent analysed is the first row of the raw tree, 2026-07-27T15:53:10.655319Z: UTC to the microsecond, under its raw key, with no age.
3. What order it’s in. The keys run A to Z, so verdict, the answer itself, is the last line.
4. Whether it agrees. Three severities share the screen: the case’s yellow priority bar, which Google’s documentation defines as Medium; the alert’s name, “MEDIUM SEVERITY ALERT”; and the agent’s severity: HIGH. Nothing says which one governs.
5. What’s missing. The conclusion runs off the panel’s edge, the side panel says the analysis was “returned” but not what it concluded, and nothing puts the two times side by side.

Six questions an analyst could ask of any verdict, and what they get today and with the card.

The same case, today and with the card

In the synthetic case CS-4133, the agent said Benign with high confidence, using data up to 13:14. At 13:57 the same key made the bucket public, and that alert, A3, joined the case. The agent hadn’t seen it. Both images are the real widget files on that case, read at 14:10 UTC.

How the card got here, from the first version through 61 variations to this one: a 62-second film.

Try it

Start on Today and find what the agent concluded, and when. Then switch to With the card, press Override verdict in Quick Actions and give a reason. Wait twenty seconds and press Refresh: the revised verdict appears, naming your override, and the Case Wall records each step.

One file, in the format the console already uses. Then four steps, in the order that pays off first.

The card is a drop-in for the widget Google ships for this action: the same action, placement, size and data binding. Once it ships with the integration, which Google now maintains, admins add it the way they add any ready-made widget. Like any HTML widget it only displays (its one documented exception, an approval link, answers a playbook step that is waiting), so the decisions are Quick Actions. The widget, its tests and a side-by-side harness are public.

With nothing else changed, it already gives the verdict its age, keeps confidence a word, and shows the whole conclusion. Here it is on Google’s example result from section 01:

Start with step 1. Its first half needs nothing new: the age comes from the timestamp the result already carries, and the action’s own Case Wall line can carry the verdict and its time, widget or not. The rest is in How to build it in Google SecOps.

The card is one answer. These eight rules are what any agent owes the console it lands in.

1. Arrive where the work already is, at whatever width the host gives you.
2. Say what you saw, and when.
3. Declare what you could not see. A missing input is part of the finding, not a footnote.
4. Let the host’s claim stand beside yours. Never silently overwrite. Never silently defer.
5. Make every claim checkable, and say where — in the host’s own objects wherever you can, and marked as leaving the case where you can’t.
6. Change nothing without an explicit human act — and say what the act will do, in both systems, before it happens. Name what cannot be undone.
7. Degrade to text. Where no widget is installed, the entry you write still carries the verdict, the confidence and the analysis time.
8. Explain your own marks. Every icon, badge and colour on your surface carries a visible label, and you never reuse a glyph the host already uses for something else.

By its documentation, they also hold for Google’s revamped case view, in preview since 26 July. An AI agent’s check found seven of the eight applied as written to a second product, Dropzone AI in ServiceNow. The details.

No analyst has used this yet: n = 0. Until one does, everything here is the card’s claim, not its result.

The test is ready: time to the right move on the hardest case, with the card and with Google’s widget, counting wrong moves. It counts against the card if analysts find the verdict as fast without it, or catch the stale verdict without it. All the limits.

The card only sees its own case. Grouping never reopens a closed case, so if the analyst trusts the verdict and closes the case, the next alert about the same key opens a new one, where the card can’t see it. That needs a check across cases, which this study didn’t design.

Some of it only a live console can settle, such as whether a widget redraws when a new result arrives: the list.

What comes next is two AIs in one case. SecOps’ own agent investigates cloud audit-log alerts, the kind that made the bucket public, so it could reach its own verdict on the new alert, with nothing to reconcile it with Wiz’s. That one is the console’s to solve.

If you work in a SOC and can spare 15 minutes, get in touch.