Technical Aspects

36 pages · sorted newest first
Technical Aspects
Mastering the technology of secure tunneling networks
Technical Aspects

Mastering the technology of secure tunneling networks

“Is a VPN just another way to use a proxy server?”…

14 Sep
Open P220
Essential knowledge about modern VPN technologies
Technical Aspects

Essential knowledge about modern VPN technologies

When evaluating remote connectivity solutions, understanding the difference between a robust corporate IPsec implementation and a consumer-grade VPN is paramount: while both offer encryption, I…

3 Sep
Open P220
How IPsec enhances your virtual private network
Technical Aspects

How IPsec enhances your virtual private network

When a field technician needs to access internal corporate resources—say, an old legacy database server running critical financial functions—while physically connected via public Wi-Fi at a client's…

1 Sep
Open P220
Comparing a VPN to an internet proxy
Technical Aspects

Comparing a VPN to an internet proxy

A Virtual Private Network (VPN) establishes an encrypted tunnel across public infrastructure to create a private, secure network connection for data transmission, whereas a proxy acts merely…

31 Aug
Open P220

Almost every technical control you will ever deploy is a trade between how much traffic you can inspect and how much traffic you must let bypass that inspection. Antivirus agents, VPN tunnels, endpoint detection, email filtering, DNS logging — each one buys you visibility at the cost of throughput, compatibility, or user patience, and every one of them has a well-documented way of being configured badly. The rest of this page is about how those controls actually work, where the seams are, and what the evidence says about whether they hold.

Antivirus testing: what an independent lab score does and does not tell you

AV-TEST awarded a maximum of 18 points per product in its March–April 2024 comparison of home-user antivirus software on Windows 11, and a product needed at least 17.5 of those points to carry the "TOP PRODUCT" designation. Of the 17 products tested, 11 reached that threshold. Read that ratio carefully: the majority of a curated field cleared the bar, which means the score separates the good from the mediocre rather than the good from the dangerous. A product sitting at 17.5 is not meaningfully less protective in daily use than one at 18; the difference is usually a fraction of a point on performance or false positives, not a missed malware family.

The mechanics matter more than the ranking. AV-TEST runs three categories — protection against malware, performance impact on the machine, and usability, meaning false alarms on legitimate software. A product can score full marks on protection and still lose ground because its real-time scanner drags on file copies or because it quarantines a line-of-business application during a definition update. For a home user, usability failures are the ones you actually feel: the installer that gets blocked, the browser extension that stops syncing, the scheduled scan that pins a laptop fan at full speed for an hour. For anyone supporting other people, a false positive costs you a support call every time, which is why the performance and usability columns deserve as much attention as the protection score.

The limit of any certification is that it is a snapshot of a lab environment against a test set. It tells you the engine caught the samples the lab chose, on the OS version the lab used, in the window the lab tested. It does not tell you how the product behaves when its cloud lookup service is unreachable, how quickly it pushes a definition for a malware family that appeared last night, or whether the vendor's own update channel has ever been compromised. Those are the failure modes that produce real incidents, and no scorecard captures them. Use the score to eliminate the bottom of the field, then choose on the things the score cannot measure: update cadence, whether the product still works offline, how loudly it nags, and whether uninstalling it leaves the network stack intact.

There is also a structural cost to running a full agent at all. Every real-time scanner sits between the filesystem and every process that touches it, so you are trading CPU and latency for detection. On a modern machine the hit is usually tolerable; on an older laptop or a machine running a heavy build pipeline it is not, and the temptation is to add exclusions until the scanner is barely watching anything. Exclusions are where most real-world antivirus deployments quietly fail — a folder excluded for performance two years ago is still excluded, and it is now where the payload lands.

How a VPN actually works, and why "VPN" is not a synonym for "the internet"

A VPN is a virtual network built on top of existing physical networks that provides a secure communications mechanism for data transmitted between networks or between nodes on the same network — that is NIST's definition in its guide to IPsec VPNs, and it is worth internalising because it corrects the most common misconception. A VPN is not a separate internet, a privacy cloak, or a magic routing layer. It is an overlay: your packets ride inside other packets across infrastructure you do not control, and the security comes from what the endpoints do to those packets, not from the fact that a tunnel exists.

IPsec is the workhorse here. NIST states plainly that the most common use of IPsec implementations is providing VPN services, which is the opposite of the popular assumption that IPsec is some optional legacy feature bolted onto a product that "really" runs on something else. IPsec operates at the network layer: it authenticates and encrypts IP packets between an endpoint and an IPsec gateway, the organisation's VPN endpoint that receives traffic sent through the tunnel. Because it works below the application, it protects everything the host sends without each application needing to know a tunnel exists. That is its great strength and the reason it survives alongside TLS-based remote access products, which operate higher in the stack and tend to be easier to expose through a browser but narrower in what they cover.

The decision a reader actually faces is not "IPsec or nothing" but where the tunnel terminates and what it is allowed to carry. A site-to-site tunnel between two offices is a different animal from a remote-access tunnel from a laptop on hotel Wi-Fi. The former is a fixed pair of gateways with known addresses and stable policy; the latter involves a device that may be compromised, on a hostile network, running whatever the user installed. Most VPN failures in practice are not cryptographic — the ciphers are fine — they are policy failures: a tunnel that authenticates too broadly, a gateway that accepts connections from anywhere because nobody restricted it, credentials reused from a breach, or a client that falls back to a weaker mode when negotiation fails. The encryption is rarely the weak link; the identity in front of it usually is.

There is a cost to tunnelling everything, and it is honest to name it. Backhauling all traffic to a central gateway adds latency, concentrates load on a single egress point, and turns that gateway into a single point of failure — when it goes down, every remote worker is offline regardless of whether their work needed internal resources at all. That pressure is exactly what pushes organisations toward split tunneling, which is where the next section lives.

Split tunneling: the efficiency gain and the inspection hole it opens

NIST strongly discourages split tunneling for IPsec VPN clients because of its security complications and risks. Split tunneling is the configuration where organisation-specific traffic goes through the VPN tunnel while other traffic takes a separate, non-VPN path — NIST describes it as a remote device establishing a non-remote connection with a system while also communicating via some other connection to a resource in an external network. The efficiency argument is real: NIST notes that split tunneling can reduce load and improve efficiency. The security argument is equally real, and it wins in most policy documents.

The reason is inspection. When all traffic traverses the tunnel, the organisation's gateway and its security stack see everything the device does — DNS queries, outbound connections, the lot. When traffic splits, only internal-resource traffic is protected through the tunnel, and everything else is excluded from it. That means the corporate firewall, the DNS filter, and the proxy see nothing of the user's other browsing. If the device is compromised while on that non-VPN path, the attacker has a channel the organisation cannot observe, and the device still holds valid credentials and a route into the internal network. NIST's telework guidance makes the same point from the defensive side: split tunneling limits an organisation's ability to inspect and protect traffic.

The trade-off is the part people get wrong in both directions. The common belief that split tunneling is always safer because less traffic goes through the VPN is backwards — NIST's guidance says it introduces security complications and can expose traffic outside the VPN, and its IPsec documentation is blunt that preventing split tunneling is more secure. But the equally common belief that split tunneling is simply forbidden everywhere is also wrong: NIST notes that the default setting allowing it is a balance between security and functionality, which is a polite way of saying vendors ship it on because users demand it and administrators often leave it on because turning it off breaks video calls and cloud apps. The practical middle ground is policy-driven split tunneling: route the specific SaaS and collaboration destinations that would choke the gateway through the local path by explicit allowlist, send everything else through the tunnel, and log the exceptions so you know what you have excluded. That is more work than a checkbox and less work than a helpdesk queue full of "my video won't load" tickets.

What you notice when split tunneling goes wrong is rarely dramatic. It is a laptop that connects to the VPN, reaches the internal file share fine, and simultaneously resolves internal hostnames through a public resolver because the DNS path split along with the traffic — a classic misconfiguration that leaks internal naming to the outside and breaks access intermittently. It is a device that looks compliant in the inventory and is quietly bypassing the proxy on every non-corporate request. Neither of these generates an alert on its own; they surface during an incident review, when someone asks how the attacker's tooling reached the internet from a managed machine and the answer is that nothing was watching that path.

What breaches actually cost, and what insurance actually covers

Insurance is a control with a hard ceiling, and the numbers from real incidents show how far below total cost that ceiling often sits. The City of New Orleans carried a cyber insurance policy limit of $3,000,000 at the time of its 2019 ransomware breach; the total estimated cost of that December 2019 attack was $7,200,000, and the city expected its insurer to cover $3,000,000 of it. The gap is not a scandal — it is how limits work — but it is the number a finance committee needs to see before it approves a recovery budget, because the difference between the two figures is money the organisation finds somewhere else.

The Pensacola figures illustrate how organisations respond. In the December 2019 ransomware incident, the ransom amount reportedly demanded from the City of Pensacola was $1,000,000. By the time of a later request for proposal for risk broker services, dated 2024, the city specified a minimum cyber liability insurance limit of $3,000,000, alongside a professional liability per-occurrence minimum limit of $1,000,000. Read together, those numbers describe a procurement posture rather than a guarantee: a floor set for the broker to beat, not an assessment that the floor is sufficient. Buyers routinely set minimums at the level their peers carry, which is how a market settles on a number that may or may not relate to any particular organisation's actual exposure.

The technical consequence is that insurance cannot be the plan for resilience. Ransomware recovery cost is dominated by things a policy may exclude or sublimit: the labour of rebuilding systems, the productivity lost while staff work around outages, the irreplaceable data that was never backed up, and the legal and notification work that follows any exposure of personal information. The controls that change the outcome are unglamorous and mostly cheap relative to the loss: offline or immutable backups tested on a schedule, network segmentation so a single compromised workstation does not reach the file server, and endpoint detection that can actually see the encryption behaviour starting. A policy limit sets the size of the cheque; the architecture determines whether you need to cash it.

The honest trade-off is that every dollar moved from insurance premium into controls is a dollar spent on a possibility rather than a certainty, and boards find that harder to approve than a policy document. The counter-argument is arithmetic: when a $3,000,000 limit meets a $7,200,000 loss, the controls are not competing with the premium — they are covering the part the premium never will.

The order of operations: what to fix first and what each step costs

Identity comes before everything else, because nearly every technical control downstream trusts it. Multifactor authentication on remote access, on email, and on administrative accounts removes the single most common initial-access path, and it costs users a few seconds per login and costs the organisation a helpdesk queue during rollout. If you are choosing between a better endpoint agent and MFA on the VPN, choose MFA; a compromised credential defeats a perfect antivirus agent because the attacker is logging in, not dropping a binary.

Second is the tunnel and its policy: restrict who can connect, from where, and to what, and decide deliberately whether split tunneling is permitted rather than inheriting whatever the client shipped with. That decision has a support cost — users will notice when their personal streaming stops working — and it is the point where a written policy saves an argument later. Third is endpoint detection with a tested response capability, not just an alert feed; an agent that reports but nobody acts on is a logging expense. Fourth is backup integrity: restoration tested, offline copy present, and a documented recovery time you have actually measured rather than assumed.

What each step costs, roughly in order of pain: MFA costs user friction and rollout effort; tunnel policy costs helpdesk tickets and occasional exceptions; endpoint detection costs licensing plus the analyst time to tune out false positives, which is the part budgets forget; backup testing costs scheduled downtime that operations teams resist. None of these is expensive in the way a breach is, but each is visible on a budget line in a way that a breach is not until it happens. That asymmetry — certain small cost now versus uncertain large cost later — is the whole reason technical controls lose funding fights, and naming it explicitly is often more useful than any product comparison.

Reading vendor claims and test results without being misled

A certification score and a vendor's marketing sentence are different kinds of evidence, and conflating them is how bad purchases happen. An AV-TEST result of 17.5 out of 18 is a bounded claim about a defined test in a defined window; "military-grade protection" is an unbounded claim about nothing. When you read a product page, look for the test name, the date, the platform, and the version tested — if any of those is missing, treat the number as decoration. The same discipline applies to VPN vendors: a claim about encryption strength tells you nothing about whether the provider logs connection metadata, whether the gateway has ever been breached, or whether the client's kill switch actually works when the tunnel drops, which is the failure that exposes traffic.

The counterintuitive claims are the ones worth checking against a primary source, because the popular belief is so often the inverse of what the guidance says. NIST's IPsec guide is a good calibration exercise: it defines a VPN as a virtual network built on existing physical networks with a secure communications mechanism, it states that IPsec implementations are most commonly used to provide VPN services, and it strongly discourages split tunneling for IPsec clients. Three sentences, three corrections to things people confidently repeat. If a vendor's documentation contradicts any of them, that is information about the vendor.

The last habit is to ask what a control does when it fails. Does the antivirus fail open and let the file through, or fail closed and lock the user out? Does the VPN client leak traffic when the tunnel drops, or block it? Does the backup job report success when the destination is full? The answers separate products that are merely certified from products that are actually operable, and they are the questions no scorecard will answer for you.