Blog/Vulnerabilities

DDoS Attacks: The Three Types and What to Do During One

How volumetric, protocol and application-layer DDoS attacks differ, why amplification works, and what to do while one is happening.

CyberWatch AI6 November 2026 · 8 min read
A crowded pedestrian crossing full of people walking in different directions

A denial-of-service attack is the only common attack type whose goal is not to take anything. Nothing is stolen, nothing is encrypted, no account is compromised. The objective is that your service stops answering, and the measure of success is your customers getting an error page.

That makes it unusual in another way: it is the attack where the useful preparation is almost entirely contractual and architectural, decided long before anything happens, and where your options once it starts are narrow. Understanding the three families and which one is hitting you is what determines whether your response helps.

The three families

VolumetricProtocolApplication layer
What it exhaustsYour bandwidthConnection state in servers and network devicesApplication and database capacity
Typical formReflected and amplified UDP floodsHalf-open connection floods, stream and state abuseFloods of requests to expensive endpoints
Volume neededVery largeModerateOften small
Blocked at your firewall?No, the pipe is already fullPartlyYes, if you can distinguish it
Hard partGetting enough upstream capacityDevice tuning and limitsTelling attack traffic from customers

Volumetric: filling the pipe

The straightforward version. Send more traffic than your internet connection can carry, and legitimate requests never arrive. The important consequence is that nothing you own can fix it: if the attack is larger than your connection, the congestion happens upstream at your provider, and your own firewall and servers may look healthy while being unreachable.

These attacks are almost always built on reflection and amplification, which works through two properties of certain internet protocols. Some services reply with far more data than the request contained, and some protocols, mostly those built on UDP, do not verify where a request came from. An attacker sends a small query to many such servers with your address forged as the sender, and each reply goes to you. The attacker's modest bandwidth becomes a flood from thousands of innocent systems, which also makes the traffic hard to attribute and impossible to block by source.

Services that have been abused this way include DNS resolvers, time synchronization servers, directory and discovery protocols, and misconfigured caching systems left open to the internet. The pattern is always the same: a UDP service, reachable from anywhere, that answers more than it is asked.

The structural fix has existed since 1998. RFC 2827, known as BCP 38, describes network ingress filtering: networks should not forward packets whose source address could not legitimately originate there. Universal deployment would eliminate the forged-sender step these attacks depend on. Adoption remains incomplete, which is why the technique still works nearly three decades later. The practical implication for you is twofold: filter your own outbound traffic so you are not part of someone else's attack, and do not run UDP services open to the internet that you do not need.

Protocol: exhausting state

Instead of volume, these attacks exploit the bookkeeping that network protocols require. The classic is a flood of connection attempts that are started and never completed, each consuming a slot in the server's table of pending connections until no slot is left for a real user.

The same thinking applies at higher layers. CVE-2023-44487, known as HTTP/2 Rapid Reset, is a good documented example: the CVE record describes a denial of service through server resource consumption, because request cancellation can be used to reset many streams quickly, and states that it was exploited in the wild from August through October 2023. CISA published an alert on it. The mechanism was not a flaw in anyone's code so much as a consequence of how concurrent streams and cancellation interact in the HTTP/2 specification, which is why mitigation arrived as a wave of updates and configuration limits across many independent server products.

What helps against this family is unglamorous: keep server and network device software current, and set explicit limits on connections, streams, request rates and timeouts per client rather than leaving defaults that assume good faith.

Application layer: looking like customers

The hardest to handle. Instead of raw volume, the attacker sends requests that are individually valid and expensive to serve. A search endpoint that queries across a large dataset, a report that aggregates records, a page that is not cached, a login attempt that triggers a password hash computation. A few thousand requests a second can exhaust a database that handles ordinary traffic comfortably.

These attacks are difficult because the requests look legitimate, which they are, and because distinguishing them from a genuine traffic surge requires knowing your normal patterns. Blocking too aggressively turns your mitigation into the outage. The OWASP cheat sheet on denial of service covers the application-side design decisions, and the limits discussed in our API security guide are the same controls viewed from the other direction.

The defenses that work are built in advance: caching so that repeated requests do not reach the database, rate limits per identity and per address with different budgets for expensive endpoints, pagination caps so one request cannot ask for everything, queueing expensive work rather than performing it synchronously, and an ability to turn off non-essential expensive features under load.

Mitigation: what it actually involves

For anything volumetric, mitigation means traffic is scrubbed before it reaches your connection. There is no on-premises version of that, which is why this is a procurement decision.

Content delivery and reverse proxy services place their own capacity in front of your site. For web applications this is the common answer, and it covers volumetric and much protocol-level attack traffic because the attack hits their network rather than yours. The caveat that catches people out is that if your origin address is discoverable, an attacker can bypass the proxy entirely, so origin addresses should be restricted to accept traffic only from the provider.

Provider-level scrubbing reroutes your traffic through filtering infrastructure during an attack. Worth asking your internet provider about before you need it, since activation is often a manual process with a lead time measured in minutes you will not want to spend.

Over-provisioning raises the bar but does not remove it, because attack capacity is cheaper than defensive capacity.

Architectural separation limits what an attack takes down. If your public website, your customer portal and your internal systems share one connection and one set of name servers, an attack on the first removes all three. Separating them is a cheap decision made early and an expensive one made late.

Three questions are worth settling with whoever hosts you, in writing, before anything happens: what DDoS protection is included rather than extra, what the activation process and expected response time are, and whether they will null-route your traffic to protect their other customers. That last one matters because for some providers the standard response to an attack on you is to discard all traffic to your address, which stops the attack and also completes it.

What to do during an attack

CISA's guidance on understanding and responding to distributed denial-of-service attacks and the NCSC's denial of service collection both emphasize the same thing: the useful actions happen with your providers, and confirming what is actually wrong comes first.

  1. Confirm it is an attack. Outages caused by a failed deployment, an expired certificate, a database problem or a legitimate traffic surge all look identical from the outside. Check whether traffic volume is genuinely abnormal and where it is arriving from. Responding to a deployment failure with DDoS mitigation wastes the time you have.
  2. Identify which family. Is your connection saturated, are connection tables full, or is the application struggling while bandwidth looks normal? This determines who can help.
  3. Contact your provider. Hosting provider, internet provider or protection service, with specific details: start time, traffic characteristics, affected services. Activate scrubbing if you have it.
  4. Preserve evidence as you go. Traffic samples, log excerpts, timings and any extortion message. You will want these afterward, and collecting them later is often impossible.
  5. Apply what you control. Enable any available protection mode on your proxy, tighten rate limits, disable expensive non-essential endpoints, increase caching aggressiveness, and block clearly abusive traffic patterns if you can identify them safely.
  6. Communicate. A status page on separate infrastructure, kept current, reduces support load substantially. Say what is affected and when you will next update. Avoid detailed technical commentary while it is ongoing.
  7. Watch for the real attack. Disruption is occasionally cover for something else, or a distraction while an intrusion proceeds elsewhere. Keep an eye on authentication logs and alerts rather than putting everyone on the outage.
  8. Do not pay extortion. It funds the next attack, marks you as a payer, and guarantees nothing. Report it instead.

Afterward, the review worth doing is narrow: what the attack exhausted first, what you could not see, how long provider engagement took, and which of your emergency measures should become permanent configuration.

Received a threat demanding payment to avoid an attack? Most such threats are never carried out. Keep the message as evidence, report it, and paste it into CyberWatch AI Scan for a free check if you are unsure whether it is a known extortion template.

Preparation that actually pays off

Almost all of the value here is in decisions made while nothing is wrong.

  • Know who to call and have the contact route tested, including out of hours.
  • Understand what your provider includes, and whether their protective response could be indistinguishable from the attack.
  • Keep your origin hidden if you sit behind a proxy, and restrict it to accept that provider's traffic only.
  • Separate your critical services so one attack cannot remove all of them.
  • Host your status page elsewhere, because telling customers what is happening from infrastructure that is down is not possible.
  • Know your normal traffic, since every application-layer decision depends on recognizing abnormal.
  • Cap expensive operations by default, which also protects you from ordinary traffic spikes.

For the network controls that surround this, see our network security guide; for the application design decisions, the API security guide; and for how resilience fits into a broader plan, the small business cybersecurity guide.

Frequently asked questions

What is a DDoS attack?

A distributed denial-of-service attack uses many systems to overwhelm a target so that legitimate users cannot reach it. The aim is unavailability rather than theft, which is what distinguishes it from most other attacks.

What is the difference between DoS and DDoS?

A denial-of-service attack comes from one source, which makes it easy to block. A distributed attack comes from many sources at once, often thousands, so blocking individual addresses does not help and the traffic can exceed your total capacity.

What is an amplification attack?

The attacker sends small requests to third-party servers with your address forged as the sender, so those servers send much larger replies to you. The attacker spends a little bandwidth and you receive a lot, with the traffic arriving from innocent systems.

Can a small website be hit by DDoS?

Yes, and small sites are easier to take down because the capacity needed is lower. Attacks against small targets are often extortion attempts, competitive sabotage or a byproduct of someone testing attack infrastructure.

Does a firewall stop DDoS?

Not a volumetric attack. If the traffic exceeds your internet connection's capacity, it is already too late by the time it reaches your firewall, and the firewall itself can become the bottleneck. Volumetric mitigation has to happen upstream of your connection.

What is an application-layer DDoS attack?

Traffic that looks like legitimate requests but targets expensive operations, such as repeated searches or report generation. The volume can be modest, which makes it hard to distinguish from a traffic spike, and it is the type most likely to evade simple defenses.

Should we pay a DDoS extortion demand?

Law enforcement advice is consistently against it. Payment marks you as willing to pay, funds further attacks and does not guarantee the traffic stops. Many extortion threats are never followed by a meaningful attack at all.

How long do DDoS attacks last?

Many are short, from minutes to a few hours, because sustaining one costs the attacker resources. Some persist for days or recur in waves, particularly when the motive is extortion and the attacker is demonstrating persistence.

What should we do first during an attack?

Confirm it is actually an attack rather than a failure or a legitimate traffic surge, then contact your hosting provider or internet provider, because the useful mitigation sits upstream of you. Preserve logs and keep a record of timings while you work.

Sources

  1. Understanding and Responding to Distributed Denial-Of-Service Attacks, US Cybersecurity and Infrastructure Security Agency
  2. Understanding and Responding to Distributed Denial-of-Service Attacks (PDF), US Cybersecurity and Infrastructure Security Agency
  3. Denial of Service (DoS) guidance, UK National Cyber Security Centre
  4. RFC 2827: Network Ingress Filtering, RFC Editor
  5. HTTP/2 Rapid Reset Vulnerability, CVE-2023-44487, US Cybersecurity and Infrastructure Security Agency
  6. CVE-2023-44487, CVE Program
  7. RFC 9113: HTTP/2, RFC Editor
  8. Denial of Service - OWASP Cheat Sheet Series, OWASP
Share this article
LinkedIn WhatsApp X
For organizations

Your people get these messages at work too.

CyberWatch AI sends your staff realistic practice attacks, trains the gaps it finds, and shows management exactly where the organization stands.