-
Person2:14:07 pm · the click
An employee opens an invoice that isn't one.
The link leads to a fake login page.
-
Agent+0.2 seconds
An AI agent blocks the page before it loads.
The reflex, as it should be.
What the employee seesThis page was blocked. It looked like a fake login page. If it's a mistake, tell us in one click.
-
Agent+40 seconds
It finds the same email in 37 inboxes and pulls it back.
Still reflex: routine, reversible and logged.
What the incident responder seesHigh Phishing email · 37 inboxes · pulled back by rule · undo available
-
Person+6 minutes
An analyst checks the agent's work, and overrides one call.
One was a real invoice. The agent was sure, and wrong. The screen shows why it decided, so the fix is one click.
What the SOC analyst seesNeeds attention 1 of 37 was pulled on the agent's own judgement, not a rule. Restore it?
-
Person+19 minutes
The agent wants to cut the laptop off the network. It waits for a person.
It stops someone's work, so it needs a yes. The evidence is on screen, so the yes takes seconds.
What the security lead seesCritical Rule R-114: a password was typed into a fake page. Isolate this laptop? Confirm or decline.
-
The law+2 hours
Did personal data leave? If so, a legal clock is already running.
Laws like NIS2 give hours, not weeks. The clock and the evidence sit side by side.
What the privacy officer seesHigh Early-warning deadline in 21 h 46 min · evidence pack ready
-
The lawNext day
The auditor signs off in minutes, not weeks.
Every step above carried its evidence, so the record already exists.
What the auditor seesResolved 6 actions · 6 with evidence attached · 2 approved by people
-
Your turn
Machines took the fast steps. People took the ones that mattered.
× every company, every dayThe agents' share keeps growing. The people's few calls must never get lost in it. Every screen above is a design decision.
Security product designer
Let AI draw the screen. Never let it guess the risk.
AI acts. People decide. I design that moment, and I measure whether it works.
Two changes, one line that must not move
Both are good, as long as risk is never guessed and people keep the calls that matter.
1 AI builds the screens
What AI may generate
Layouts, charts, summaries, colours. They can change every time.
Three generated screens for one alert. Only the label stays the same.
What it must never guess
The words and colours that say how serious something is. Fixed, and the same on every screen.
- Critical follows written rules, can't be dimmed
- High needs attention soon
- Needs attention no ranking yet
- Resolved with the evidence attached
2 AI agents take actions
AI agents are becoming security's reflex. An agent can earn the right to act alone, one action at a time, like a promotion. It never earns the right to guess how serious something is. The design job is deciding what reaches a person, and when. One afternoon, made up but typical:
3 Across the whole of a company's security
8 parts of a company × 9 decisions = 72 moments a person decides
Each dot is one screen where a person makes a call. Filled: the afternoon you just read. Hollow: the same moment elsewhere, still to be designed well.
| Parts of a company ↓ · Decisions → | Author | Approve | Override | Except | Classify | Respond | Notify | Attest | Interrupt |
|---|---|---|---|---|---|---|---|---|---|
| Endpointlaptops, phones and servers | |||||||||
| Accesswho can reach which apps and websites | |||||||||
| Datawhere sensitive data goes | |||||||||
| Identitywho each person and AI agent is | |||||||||
| Exposureweak spots, before attackers find them | |||||||||
| Detectionspotting and stopping attacks | |||||||||
| AIthe company's own AI agents and models | |||||||||
| Governanceproving it to auditors, boards and regulators |
Someone disagrees with the AI about an attack alert.
The work, with its limits
Each links to its method, including results that went against me.
- 25 → 65 → 90% on-system
Running a design system an AI writes against
A component library on its own — ours — got generated UI about a quarter of the way onto the system. Writing the rules machine-readably added forty points; testing those rules added the last twenty-five, alongside more components acquiring docs. I don’t split that last twenty-five. Correction per screen, before and after: ~2 hrs → 15 min.
Enterprise security platform · 2022 – 2026my own judgement, approximate · no control and no second rater · no count of the screens · prompt not held constant - 93.2% to 100%
A rule without a value is not a rule
I published a claim about why AI-generated UI drifts off-system. Then I built a controlled reproduction to test it. The claim was wrong. What moved the number was naming the value.
Field test 01 · August 2026five components, 74 properties, three runs per condition, one model · not a benchmark - eight rules
The agent is done. The case isn’t.
Google SecOps ships a widget for Wiz’s AI verdict. This drop-in replacement gives the verdict its age from today’s fields. The rules behind it are written to carry to other consoles.
Verdict card for security consoles · Self-initiated · 2026drop-in widget, open source · synthetic console, no tenant · no analyst has used this yet: n = 0 32 locations, one view
I put every location on a row, every check in a column, and the health of each pair where they meet. The whole estate now reads in one pass, and you fix a problem from the same place you spot it.
VMware NSX · 8 weeks · 2021–22shipped in NSX · both concepts: unmoderated usability testing in Maze, 43 participantsTwelve selected, twelve different values
Select twelve servers and change one setting. The twelve don’t agree with each other. What should the field show?
Bulk edit · 2026not shipped · n = 0Three per cent
Buttons in Vela, the design system I build, shrank three per cent when you pressed them. One line of CSS, copied from somewhere, never questioned. It was throwing clicks away. A press no longer moves its own target.
Vela · 2026shipped · npm · mouse users only · measured on the prototype in Chromium and WebKit, on the shipped button in Chromium
Work with me
For founders and CTOs of security vendors. Start with a fixed-price pilot, then keep going if it works.
Pune, your time zone · Security UX since 2020 · Patent pending, US 2025/0036264
Understand, redesign, prove
- Understand: the five issues that matter most, ranked.
- Redesign: a clickable prototype.
- Prove and hand over: three to five real users try the prototype on real tasks; dev-ready spec and design tokens your tooling can read.
Length and fixed price set after a first call, to fit your scope · 50% on acceptance, 50% on handover
How the pilot worksWhat I'm learning in the open
Unfinished questions, shared early. Nothing here is for sale yet.
Where security software is heading
Forecasts, with how sure I am. I'll score them in public as they resolve.
By 2029, AI agents do most of the work inside security products. The scarce skill is designing where a person reviews, overrides and audits them.
Hypothesis · being tested Evidence so far
Two forces are already pushing it there.
1 The areas are merging
In the grid above, Access, Data and Identity are separate rows. Inside real products they overlap, and the hardest calls sit where they do. No single product owns them, so no single team designs them.
2 Laws now set the clock
- NIS2 (EU)24 h · 72 h · 1 month
Early warning, notification, final report. Applies as each country writes it into law. source
- EU Cyber Resilience Act24 h
Makers warn early on an actively exploited weakness. source
- India DPDP
Consent managers begin: one year after notification on 13 Nov 2025. source
- India DPDP
Most duties, including breach reporting: eighteen months after notification on 13 Nov 2025.
What moved this fortnight
A fortnightly brief on security product design across the whole estate. Three recent items below. Issue 1 is out on LinkedIn.
Microsoft previews an integrated SOC in Defender: XDR, SIEM and AI in one portal, so one case can mix native and ingested evidence.
Wiz extends its partner network with MCP-powered agent integrations: a product's answer to another company's agent is now a design surface.
Sophos: 46% of MSPs say customers rely on them to act as their CISO, and 55% still do part of their security reporting by hand.