Use It Where It Works

Five things to do this week

Tick them off as you go. Ticks stay on this device only.

Run the three questions (below) on one AI use you already have.
List the detection rules that have never fired. Then decide whether that is good news.
Write down your "never automatic" list for anything AI can act on: account recovery, payments, access changes.
Ask the four questions of anything AI helped build, before it touches real data: what can it reach, who approves, can we recover, how far does it spread.
Pick one policy. Have AI turn it into steps with a clause citation on every line. Have someone who did not write it follow the steps once.

The three questions

Ask them of any AI proposal. They do not change between workflows; the answers do.

1 · Does wellWhich parts does AI genuinely do well here?

2 · Data firstWhat has to be true of the data before it can?

3 · Process around itWhat has to sit around it?

	1 · Does well	2 · Data first	3 · Process around it
Security operations	Volume: alerts, intelligence, drafting detections	Detections that actually fire on what is logged	Route by confidence; give it its own identity and limits
Software development	Writing code fast	A clear, corrected specification	People who understand it build what touches real data
Security posture	Fine print into clear process	Knowing what you actually do and accept	Cited sources; a named owner signs off

Questions to take back to your team

Tap a workflow to open it.

Security operations

Where do analysts spend hours reading, not deciding?
Which decisions must a named person be able to explain later?
Which of our detection rules have never fired at all - and is that good news?
Who samples the automatic lane each week to check it is still right?
Is the reviewer given the evidence, or only the model's conclusion?
Who is on the other end of an escalation at 2am?

Any AI that can take an action

What does our assistant accept as proof of identity?
Can the model call a sensitive action directly, or only request it?
Would we notice the same request a thousand times in an hour?
Which actions are on our "never automatic" list - and is it written down?

Software development

Is AI-written code labelled, so we know what to review harder?
Does a second person or tool review what the AI wrote?
How many steps does this workflow chain together, and what is the measured accuracy of each?
What can it reach? Development or production? Read or write? Can it run a migration, delete a table, deploy?
Who approves? Before it acts, or after? Would we give a new joiner that access unsupervised?
Can we recover? Has anyone restored from a backup recently? Can the credentials that write the data also delete the backups?
How far does it spread? Private data + the ability to act outward + reading content it did not write. Any two are manageable; all three is the difficult case.

Across the board

Which of our runbooks are current enough to hand to an AI?
What would we lose if two key people left tomorrow?

Checks you can run

Small, concrete, and none of them need AI.

Apps built with AI

Three checks before it touches real data

Is access control switched on for every database table?
Search the published app for anything that looks like a secret key.
Log in as one user and request another user's record.

Third-party access

What would have caught Denmark

Strong sign-in on every partner account with access to sensitive data.
A limit on lookups per account, and an alert when it is exceeded.
A baseline of each partner's normal use, reviewed daily.

Policy and compliance work with AI

Four checks so it holds up

Every framework mapping cites the clause on both sides.
Someone who did not write a procedure follows it once.
A person confirms each gap is real, not a wording difference.
Every statement traces to a record of something that actually happened.

The AI drafts from the source documents you give it, citing the clause for every line. A named person chooses the sources, checks the citations, decides what the organisation accepts, and signs off.

Rules of thumb

AI is fluent well before it is correct.
It works on expertise that can be written down - so write it down.
Point AI at the detection gap, not only at the alert queue.
Decide in advance which lane each kind of output goes to.
Give anything that acts its own scoped identity, the way you would a new joiner.
Correct the specification, not the output.
Guardrails do not fix it. People route around controls that slow them down.
The AI drafts. A named person decides and signs.
Own the system, rent the model.
Get clear before you automate.

Sources

Every figure in the talk, so you can check it.

Picus Security, Blue Report 2026 (11 Aug 2026) - 58% of simulated attacks logged, 14% alerted, 37% of post-compromise actions blocked.
Splunk, State of Security 2025 - 74% rate detection engineering the most important future SOC skill.
CPR Administration, Denmark (5 Oct 2026, in Danish) - unauthorised access to about 8.8 million people's records through a company's lawful access.
Krebs on Security (1 Jun 2026) - an AI support bot talked into handing over accounts.
CVE-2025-48757 (NVD; the record is marked disputed by the vendor) and the researcher's statement - 303 exposed endpoints across 170 AI-built apps.
Spracklen et al., "We Have a Package for You!" (USENIX Security 2025) - 19.7% of AI-recommended packages did not exist; and how attackers register them.

"Small errors multiply": 95% per step is 77% over 5 steps and 60% over 10. Arithmetic, not a survey.