Five seconds to know.
One click to act safely.
A cybersecurity hiring brief, shared with me as a practice exercise. I set myself a three-day timebox to design an investigation side panel and a presentation that could stand on its own.
What I was asked to design.
A security operations center analyst receives a high-risk alert. My task was the panel used to investigate it, within an existing dashboard.
Daniel, SOC analyst.
Many tools open, many alerts, little time. He needs a quick first read before deciding where to investigate.
Risk 95.
A Finance employee downloaded 50 GB of customer data at 2 AM, from an unrecognized device.
Four challenges, one flow.
Hierarchy, explainable AI, actions with guardrails, and drill-down. One complete flow that reviewers could follow without me.
| The brief asked for | In other words | Where it lives |
|---|---|---|
| Hierarchy and AI summary | Who, what, how severe, and what next. | 01 · The first read |
| Explainable AI | Show what makes this suspicious. | 02 · Why 95 |
| Actions with guardrails | Easy to reach, hard to get wrong. | 05 · Block · 06 · Suspend |
| Drill-down | Explore files and devices without losing context. | 03 · Files · 04 · Device |
| A complete flow | From opening the alert to containment and escalation. | 01 → 07 |
Four criteria before any screen.
I translated the brief into criteria I could check each design against.
5-second read
Who, what, how severe and what to do, before scrolling.
Trust
Each factor next to its baseline, with counter-evidence visible.
Safe action
Narrow and reversible actions are quick. Disruptive actions show their impact first.
Keep context
Drill-down preserves the alert, score and next step.
Behavior before pixels. First what Daniel needs to understand, then how it looks.
Three insights shaped the panel.
AI-assisted desk research across industry reports, studies and product references informed these choices. This was not a user-interview study.
The score is a claim. The baseline is evidence.
A 95 is hard to assess in isolation. Comparing each factor with normal activity makes the reasoning inspectable.
It may not be her at all.
An unfamiliar device keeps account takeover on the table. The AI should weigh explanations, not judge a person.
Friction should match the impact.
Identical confirmations encourage automatic clicks. The action’s scope and consequences need to be clear.
False positives were the top detection challenge for 73% of respondents in the 2025 SANS Detection & Response Survey, according to Stamus Networks’ survey summary ↗.
Product references: CrowdStrike · Microsoft · SentinelOne · Varonis · Wiz · Island · Wonderful
Why 95, factor by factor.
A demo score breakdown makes the hierarchy concrete: the strongest signals first, each next to a baseline.
Illustrative contributions total 95 points, not percentages. Her team’s highest day in 90 days was 2.1 GB, about 24 times less than this event.
The first read
Title, score, verdict, next step and counter-evidence.
The case
Factors, baselines, hypotheses and response plan.
The proof
Files, device details and logs beside the panel.
Three options each. One direction chosen.
I compared low-fidelity options against the four criteria. Here are the selected structures and the final screens they informed.


Verdict first
The verdict and next step come first. Entity-first answered who before what; score tiles gave a number without an explanation.


Factors vs. baseline
Each claim sits next to a baseline, with counter-evidence visible too. A score followed by a paragraph would still hide the reasoning.


A response plan
The sequence explains what to do next, and friction grows with impact. Equal buttons would flatten those differences; automatic containment would remove the human decision.


Detail beside the panel
The verdict and response plan stay visible during drill-down. Tabs would hide the actions; a separate page would lose the alert.
I rejected the first visual draft as too generic. The structure stayed; a dark verdict card, stronger type hierarchy and purposeful color made the priorities clearer.
One alert, one complete response.
Seven screens tell a simulated night from 02:51 to 02:56. The timestamps illustrate the intended flow, not a tested completion time.

Opens the alert
The verdict and counter-evidence, in one view.

Why 95
The volume compared with her normal activity.

The files
96% customer PII, from one folder.

The device
A known iPhone approved an unfamiliar device.

Block
A narrow action with a 30-second undo.

Suspend
The impact, a required reason, and an owner.

Contained
Escalate with the full investigation record.
Select any screen to view it full size.
Answer first.
One dark card holds the verdict, score and next step. Counter-evidence sits just below to help the analyst question the AI’s conclusion.
Go deeper, keep the verdict.
Files and device details open beside the panel, preserving the alert and response plan.
Friction matches the impact.
Block has a short undo window. Suspend requires reviewing the impact and entering a reason. Escalation remains a human decision.
Evidence that sharpens the next question.
The unknown laptop sits beside her managed devices. A known iPhone approved the sign-in, but that alone cannot establish who used the laptop. The design keeps uncertainty visible while supporting a narrow first action.
- No security agent and no IT record are explicit.
- Known devices provide a comparison.
- Blocking updates the same device state throughout the panel.

What I assumed, and how I’d check.
These dependencies would need validation with the product and security teams before implementation.
Model explainability
Confirm which factors and baselines the engine can reliably expose.
Permissions and policy
Check containment permissions, escalation ownership and employee-impact decisions.
Reversibility
Verify which actions can be undone and what a 30-second window actually guarantees.
What I’d measure
- Time from opening an alert to an informed decision.
- How often a closed alert is reopened.
- How often Undo is used, and why.
AI sped it up. I owned the decisions.
Where AI helped. Research synthesis, first drafts of the framing and hypotheses, evidence-model options, edge cases, and screen construction in Figma.
What I owned. Reviewing those drafts, choosing the four design directions, setting the visual direction, rejecting the generic first pass, and approving the final screens.
What I’d explore next. Pseudonymized triage until escalation, and a second flow for closing the alert as legitimate when IT confirms the device.