In 2024 I wrote Security for Web3 Organizations, a starter guide for small teams with nobody owning security: name someone responsible, keep lists of users, assets and software, harden the obvious, write the few procedures that matter, and run the recurring checks. It was a checklist for organisations that had no one to hand the problem to.
This post is what that thinking became once the question changed from what a team should do to what the person who owns security should do first. The lists are still here; they grew into the inventory at the centre of the first ninety days. The biggest change since then is AI. Much of the recurring work that guide expected a small team to do by hand (the reviews, the checks, keeping the lists current) can now be handed to agents, which changes what a small team can cover and where its people should spend their time.
If you are about to take a security leadership role, this is the plan I would want handed to me on day one. If you are hiring for one, this is roughly the plan I would expect a good candidate to bring — and the last section is what I would listen for.
The stance underneath it
The plan assumes one stance, and everything else depends on it. Security does not block. A function that blocks gets routed around, and a security function that is routed around is worse than none, because it produces the appearance of coverage while the actual decisions happen somewhere you are not. My job is to find risk, size it, put it in front of the person who owns the outcome, and write down what they decide, with an owner and a date. There are two exceptions: anything that cannot be undone, where I want a named human making the call and will say no, and commitments the company has already made, which I enforce as facts. The inventory in week two, the choice of tabletop scenarios and the last item on the deferral list all follow from that.
The arithmetic
A security team at almost any modern company owes the organisation somewhere around twenty distinct services (incident response, detection, design review, vulnerability management, vendor review, questionnaires, bounty triage, identity governance, awareness, exception handling) plus a recurring calendar underneath them that runs to roughly seventy tasks once you write it down: daily alert queues, weekly burn-down, monthly access and posture reviews, quarterly recertifications, an annual audit and pentest cycle.
That number is set by your systems and your obligations. It is not set by how many people you have.
So a team of three or four faces a workload defined by the size of the estate, not by the size of the team. The team ends up spending all of its time in the queue and none of it thinking, and within a year you have a ticket desk with a security label on it.
Hiring does not close that gap, because every new system adds queue faster than a hire subtracts it. Outsourcing helps, but less than it looks. A managed detection provider is the right answer for coverage at three in the morning, and at this size I would buy one. What it cannot sell you is judgment: the provider does not know your business well enough to make the calls that matter, so every alert it escalates still lands on your team. Working harder is the black hole.
The only thing I have found that closes the gap is automating the toil between the outsourced eyes and the in-house decisions, and keeping the judgment. In 2026 that means agents doing the recurring work end to end, inside a hard boundary on what they are permitted to do. The boundary moves in one direction only, and every move is earned:
Autonomous for work that is logged, reversible and read-only.
Supervised where the machine does the work and a human clicks.
Human-led for incidents, threat models, and anything in the irreversible class.
The governance matters more than the automation. An agent with production credentials is a privileged identity, and if you would not give a contractor that access without a written scope, you should not give it to a process either.
That is a whole post of its own and I will write it separately. For now: if you are planning a small security function on the assumption that it will scale by working harder, the plan is already wrong.
The first ninety days
I do not think much of ninety-day plans built entirely out of reading and listening. Documents and interviews tell you what an organisation intends. I want to know what it does under pressure, and the fastest way to find that out is to apply some.
Evidence first, then pressure — and the evidence work never stops once the pressure starts.
Month one: evidence, and keeping the work moving
Nothing stops while you read. If there is anyone already doing security work, your first job is to understand what is in flight and keep it moving rather than interrupt it. Leadership will also have one or two things they want dealt with immediately; handle those in month one rather than queuing them behind your assessment. You are building the credit you will spend later.
Then gather evidence:
Every incident and near-miss of the last six to twelve months, read end to end. Skip the summaries and read the threads, the calls, the post-mortems that never got written.
Every security questionnaire and every deal that stalled with security named as a factor. This is a two-day exercise with sales, and it is the only honest input to the certification question later.
Interviews across the company, not just engineering, with one question doing most of the work: what stops you from doing your job? People answer that one honestly. The answer tells you where process gets skipped, and where process gets skipped is where incidents come from.
An inventory of what cannot be undone. Who can move money. Who holds production. Who can change DNS, the registrar and the email configuration. Who can reset an account or an MFA factor. This is a list, not a project, and it should exist inside two weeks.
Month two: pressure, and the evaluation
Run tabletop exercises, with the people who would actually be in the room. Three is enough, and I would pick them by what cannot be undone rather than by what is likely:
An irreversible action gets approved that should not have been.
A credential compromise reaches the release or deploy path.
A customer or partner tells you before you know.
Nobody runs that third one, and it is the one that tells you the most.
Participate; do not grade. Running exercises as an audit in your first month, before you have earned anything, is the fastest way to lose the room.
The real output is the ownership map: who owns the registrar, the CDN, the DNS, the signing keys, the customer comms channel, the third-party dependency nobody has thought about since it was integrated. Half your roadmap gets written from the questions nobody in the room could answer. A two-hour exercise surfaces a gap list that would otherwise take you two quarters to assemble.
State the limitation honestly, though: tabletops test response and are blind to posture. They will tell you nothing about an unpatched fleet or secrets sitting in a repository. Which is why the evidence work from month one keeps running underneath them rather than stopping when they start.
That posture view is the other half, and it starts with an inventory of people, process and technology: who holds which access and which knowledge, which processes actually run and who owns each one, and every system, vendor and integration you are responsible for. Those are the three buckets I have used to classify every incident I have looked at for years, because they are exhaustive and they map to different kinds of fix. Rate every entry and order the whole list by risk. The month-one list of what cannot be undone becomes its top tier.
That prioritised inventory is the most important thing you produce in the first ninety days. Everything else in the plan is scoped against it, and you cannot protect what is not on it. The tabletop gaps tell you what is missing; the inventory’s ordering tells you what to fix first. Merged, they are your roadmap.
Also in month two: start a monthly risk forum with senior leadership. Do not open it by asking what the company’s risk appetite is; that question produces a sentence nobody can act on. Appetite is established by decisions. Bring real acceptance decisions to every session, watch what gets signed and what gets pushed back, then write the pattern down and read it back for ratification.
Month three: commit
The priority plan written.
An incident response plan that says who decides what, built from what the tabletops exposed rather than from a template. Inside it go the two things Two things kill a company argues decide the seventy-two hours after the bad day. One is a disclosure standard (what you tell customers, partners and regulators, how fast, and who signs it off), agreed now while nobody is panicking. The other is a named human that customers and partners can escalate to, published where they will find it. The tabletop where a customer or partner tells you before you know is the input to both. None of this prevents anything. It determines whether the next incident is survivable, and it is the work that cannot be done after the fact.
The certification decision made on the evidence you gathered in month one rather than on reflex. Your first hire requisition open.
And a first draft of the risk appetite, written from the decisions the forum has made so far, read back to leadership, and explicitly labelled a draft. Two sessions are not enough to call anything a pattern.
The other nine months
Month three hands you an ordering, and most of the remaining nine months are spent working down it from the top. Three things run underneath it regardless of what it contains. The operating cadence (access reviews, named asset owners, evidence captured as a by-product of the work, a risk register someone actually maintains) gets built whether or not the company decides to certify. The recurring work moves behind the boundary described above, because the arithmetic does not improve on its own. And the monthly risk forum keeps running: appetite is established by decisions, so the month-three draft gets ratified around month six, once there are enough of them to read a pattern from, and revised again by the decisions that follow.
What I would defer, and why
I would judge a candidate on this list harder than on anything else. The judgment in this job is mostly about what does not happen in year one.
A 24/7 in-house operations centre. Not achievable at small headcount and dishonest to promise. Buy overnight eyes from a managed detection provider, keep the escalation decisions in-house, back it with an incident response retainer, and publish exactly where that line sits rather than letting people assume more coverage than exists.
Full coverage of everything. Coverage follows value. The long tail gets liveness and gross anomaly alerting until it earns more.
A formal security champions programme. Everyone puts this in a twelve-month plan and almost nobody staffs it. Below a certain size you can just talk to every engineer who matters.
Rewriting the policy suite. Policies nobody reads are audit theatre. Write policy when a certification, a customer or a real decision requires it, and not before.
A tooling consolidation project. Tempting, visible, and it produces no risk reduction in a year where the gaps are elsewhere.
Anything that makes you a blocker in week three. Including your own best idea. Inserting yourself into the path of everything that ships, before anyone trusts you, is how a security function gets permanently routed around. That is much harder to undo than the risk you were reducing.
On certifications
Certifications do two things. They create trust with customers and stop security dragging on the sales cycle. And they impose an operating cadence (access reviews, named asset owners, evidence trails, a risk register that actually gets touched) that an organisation would otherwise have to invent, sell internally and enforce by force of personality.
The second is the underrated one, and you can have it without paying for the certificate.
So I separate them. The structure gets built either way: map the work to a recognised backbone, and capture evidence continuously as a by-product rather than reconstructing it six weeks before fieldwork. I spent a decade maintaining certifications rather than winning them once, and that distinction is the entire difference between a calm audit and a terrible quarter.
Whether you certify is a business decision, not a security one. Bring the evidence, the cost, the timeline and a recommendation, and let the executive team decide. Done in that order, certification becomes a three-to-six-month sprint whenever the business wants it, rather than a twelve-month rebuild. The option stays cheap.
If you are the one hiring
Four things I would listen for, in roughly this order.
Can they tell you what they would not do? A plan with no deferral list is a plan that has not been costed. Anyone can list everything that ought to happen.
Do they know where they stop saying no? If the answer is “security should be able to block,” they will be routed around within a year. If the answer is “I never block,” they have not thought about the irreversible class.
Do they plan for the arithmetic? Ask how a team of three covers an estate that generates twenty services and seventy recurring tasks. If the answer is headcount, the answer is wrong.
Do they treat the first ninety days as evidence-gathering or as opinion-forming? The candidates worth hiring want to find out what your company actually does under pressure. The ones to avoid arrive already certain.
And one thing to offer rather than test: tell them who owns IT, who holds incident authority, and what has actually gone wrong in the last year. Every plan they wrote before joining was built on assumptions, and the faster you correct those, the faster the plan becomes real.
