Security is a stack of controls that each fail differently, and the useful question is never "am I protected" but "which control catches this, and what happens when it doesn't." Antivirus, patching, a VPN, backups, insurance and incident response are not interchangeable layers of the same thing — they cover different failure modes at wildly different costs, and getting the order right matters more than getting the best product in any one category. The rest of this page walks through that stack in the order a defender actually has to build it.
Endpoint protection: what a lab score does and doesn't tell you
AV-TEST's March–April 2024 comparison of home-user antivirus products on Windows 11 tested 17 products, and 11 of them earned the "TOP PRODUCT" designation by scoring at least 17.5 out of a possible 18 points. That tells you the ceiling is crowded: the difference between a top-scoring product and the one ranked just below it is small, and no consumer product in that field is dramatically worse at detecting known malware than its neighbours. What the score compresses is protection, performance and usability into one number, and the differences that bite in practice — how aggressively the product quarantines a legitimate installer, how much it slows a build, whether its ransomware rollback actually restores files — are exactly the things the aggregate hides.
So treat the score as a floor, not a ranking. Pick a product that clears the TOP PRODUCT bar, then judge it on the things a lab can't measure for you: whether it interferes with your work, whether it has a credible behavioural detection layer rather than signature matching alone, and whether its ransomware protection keeps versioned copies somewhere the malware can't reach. Signature-only protection is close to useless against a targeted intrusion, because the attacker tests against the major engines before deploying. Behavioural blocking is what catches the novel binary; the signature engine catches the commodity stuff and stops it consuming your attention.
The cost is real. Endpoint agents on every machine are a management burden, they generate alerts someone has to triage, and on underpowered hardware they make everything slower. A small organisation with no one to read alerts is often better served by a product with strong defaults and few knobs than by an enterprise console nobody logs into. The wrong choice is buying the top of the range and leaving it unmanaged — a console full of ignored detections is worse than no console, because it creates the belief that someone is watching.
Patching and identity: the two controls that stop most of what reaches you
Most intrusions that succeed do so through an unpatched exposed service or a credential that was reused, phished or bought. That is not a claim about any particular breach; it is the shape of the problem, and it's why patch cadence and identity hygiene outrank almost everything else you can buy. The sequence that works: inventory what's actually exposed to the internet, patch the internet-facing things first and fastest, then work inward. A monthly patch cycle for workstations and a much shorter one for anything with a public IP is the standard shape, and the reason to shorten it is that exploitation of a freshly published vulnerability now typically follows disclosure closely enough that "next patch Tuesday" is too late for exposed systems.
Identity is the other half, and it's where the cheap wins are largest. Multifactor authentication on email, on the VPN, on anything administrative, removes the value of a stolen password entirely. Hardware security keys resist phishing in a way that push notifications and SMS codes do not, because the key checks the origin of the site asking. Password managers exist so that no human is choosing or remembering credentials, which is the only realistic route to unique passwords everywhere. The trade-off is administrative: MFA generates lockout tickets, hardware keys get lost, and a poorly rolled-out MFA programme will produce shadow accounts and workarounds that quietly undo it. Roll it out with a break-glass path and expect the first month to be annoying.
What this approach costs is attention and a little friction, and what it does not do is stop an attacker who already has valid credentials and a session token. That is why the next layer exists.
Remote access and the VPN question
A VPN is not a privacy product and it is not a synonym for the public internet. NIST, in its guide to IPsec VPNs, defines it as a virtual network built on top of existing physical networks that provides a secure communications mechanism for data and information transmitted between networks or between different nodes on the same network. IPsec is the workhorse: NIST states that the most common use of IPsec implementations is providing VPN services, with the organisation's IPsec gateway as the endpoint that receives tunnelled traffic. If you are building remote access for staff, that's the technology underneath almost everything you'll evaluate.
The configuration decision that causes the most argument is split tunneling — letting a remote device send organisation-specific traffic through the tunnel while other traffic takes a separate, non-VPN path. The folk wisdom is that it's safer because less traffic traverses the VPN. NIST says the opposite: split tunneling introduces security complications and risks, and its IPsec VPN guidance strongly discourages it for IPsec VPN clients, noting that preventing it is more secure while the default setting that allows it is a balance between security and functionality. Its telework guidance is precise about the mechanism — split tunneling protects only internal-resource traffic through the VPN and excludes other communications from the tunnel, which limits the organisation's ability to inspect and protect that traffic.
So the honest trade-off is performance and cost against visibility. Full tunneling pushes all traffic through your gateway, which means you can inspect it, filter it and log it, and which also means your gateway and your internet breakout become a capacity and single-point-of-failure problem. Split tunneling reduces load and improves efficiency, and it hands the untunnelled portion of the user's traffic to whatever protection their local network and device provide. For a small organisation with a cloud-first footprint, a zero-trust access broker in front of individual applications often beats a full-tunnel VPN entirely, because it exposes one app rather than the network. For a organisation with on-premises systems and a compliance requirement to inspect egress, full tunnel is the defensible default and the bandwidth bill is the price of it.
Backups, and why insurance is not a control
Backups are the only control that reliably survives a successful ransomware intrusion, and the only thing that makes them reliable is that at least one copy is offline or immutable and that you have actually restored from it. The failure mode is consistent: the backups exist, they're on a network share the compromised account can reach, and the attacker encrypts them along with everything else. The sequence that works is the 3-2-1 shape — multiple copies, on different media, with one off-site — plus a tested restore, because an untested backup is a hypothesis. Test the restore on a schedule you can defend, and time it: the number that matters in an incident is how long a full recovery takes, not how many terabytes you've stored.
Cyber insurance sits alongside this as a financial instrument, not a security one, and the numbers show why it can't be depended on to cover an incident. The City of New Orleans carried a $3,000,000 cyber policy limit at the time of its December 2019 ransomware breach; the total estimated cost of that attack was $7,200,000, and the city expected the insurer to cover $3,000,000 of it. That gap is the whole lesson. In the same period, the City of Pensacola reportedly faced a ransom demand of $1,000,000 in its December 2019 incident. By 2024, Pensacola's RFP for risk broker services specified a minimum cyber liability limit of $3,000,000, alongside a professional liability minimum of $1,000,000 per occurrence — which is a fair picture of what a mid-sized public entity is now expected to carry, and of how the two policies interlock.
Buy the policy, read the sublimits, and assume you will fund the difference. Insurers now ask for evidence of MFA, backups and endpoint protection before they'll quote, which means the security work is partly a precondition for the insurance rather than a substitute for it. The wrong choice is treating the policy limit as your recovery budget; the right one is treating it as the part of the bill someone else pays.
Detection, response and the plan you write before you need it
You will not detect a competent intrusion with antivirus alone. What catches it is telemetry that's centralised enough to correlate — authentication logs, endpoint process events, network flows — and someone or something looking at it. For most organisations that means shipping logs to a managed detection provider rather than building a security operations centre, because the cost of a 24/7 human rota is prohibitive and the value of a tool nobody watches is zero. The decision to make is what "good" looks like on a Tuesday: how long between an alert firing and a human acknowledging it, and who has authority to pull the plug on a system at 3am.
Response is where preparation pays and improvisation fails. The plan that works is short, names people rather than roles where possible, and covers the first hour: who declares an incident, who talks to the insurer, who talks to legal, who handles communications, and what the isolation procedure is for a compromised endpoint. Keep an offline copy, because the plan stored on the file server is unavailable exactly when you need it. Run it once as a tabletop exercise and you will find the gaps — the insurer's notification deadline, the fact that nobody has the gateway admin credentials, the backup restore that takes longer than the business can tolerate.
The honest cost here is that incident response capability is expensive and mostly idle, and the return is invisible until the day it isn't. A small organisation's realistic position is a retainer with an incident response firm, a written one-page plan and a tested restore, rather than a detection stack it can't operate.
What the whole stack costs, and how to sequence it
Build in this order: identity first, patching second, backups third, endpoint protection fourth, remote access architecture fifth, detection and response last — because each earlier layer reduces the volume of work the later ones have to do, and because the early layers are cheap relative to their effect. MFA and a password manager cost little and remove entire categories of attack. A patch cadence costs process discipline. Tested offline backups cost storage and a recurring calendar entry. Everything after that is progressively more expensive per unit of risk removed.
| Layer | What it catches | What it costs | Where it fails |
| MFA and password manager | Stolen and reused credentials, phishing | Licensing plus user friction | Session token theft, MFA fatigue, poor rollout |
| Patching | Exploitation of known vulnerabilities | Change management time and testing | Exposed systems patched on a slow cycle |
| Backups | Ransomware, destructive attacks, deletion | Storage, plus restore testing time | Online copies reachable by the attacker; untested restores |
| Endpoint protection | Commodity malware, some novel binaries | Agent overhead and alert triage | Unmanaged consoles, signature-only detection |
| VPN / access architecture | Exposure of internal services to the internet | Gateway capacity, or broker licensing | Split tunneling leaving traffic uninspected; flat internal networks |
| Insurance | Part of the financial loss | Premium and underwriting evidence | Limits below actual cost, as in New Orleans |
The judgement call that matters most is not which product to buy but how much to concentrate on prevention versus recovery. Prevention fails silently and eventually; recovery is what determines whether a bad week becomes an existential one. A organisation that has MFA everywhere, patches its exposed systems quickly, keeps an immutable backup it has restored from, and carries insurance sized against a realistic loss estimate is in a far stronger position than one that has bought the best-scoring endpoint product and left the rest alone. The 11 products that cleared AV-TEST's TOP PRODUCT bar in March–April 2024 are all broadly fine; none of them will save you from a valid credential and an unpatched gateway.