Cloudflare’s account of the 5.6 Tbps attack it mitigated in October 2024, published in its Q4 2024 DDoS threat report, carried a detail that got far less coverage than the headline number: the flood lasted roughly 80 seconds. A record-breaking volumetric event, over before most on-call engineers would have finished reading the first alert. That one data point explains most of what is useful in DDoS attack statistics trends, and most of what misleads buyers who read the quarterly reports as though they were a single shared scoreboard.
They are not a shared scoreboard. Each report is a view from one vendor’s sensors, counted by that vendor’s rules, over that vendor’s customers. Read properly, the numbers still tell you something worth acting on. Read as a league table, they will sell you the wrong protection.
Why DDoS attack statistics trends disagree so much between reports
Start with the obvious thing nobody states in the executive summary: a threat report is telemetry from a customer base. A content delivery network sees what its web customers see, which is HTTP request floods against web properties. A transit carrier sees what crosses its backbone, which is volumetric UDP and reflection traffic. A scrubbing provider sees only the events its customers routed to it, which excludes everything the ISP absorbed upstream and everything that never reached a protected prefix.
So a 500 percent rise in one report and a decline in another over the same quarter are not two readings of the same threat. They are measurements of two different populations. Comparing them is a category error, and it is the most common one in boardroom slide decks.
Counting rules move the totals just as much. Is a four-hour campaign that pauses twice counted as one attack or three? Is a multi-vector event with a UDP flood, a DNS query flood and an HTTP flood counted once, or once per vector? NETSCOUT’s DDoS Threat Intelligence Report and Cloudflare’s quarterly reports both publish their methodology, and the methodologies are not the same. When either vendor adjusts detection sensitivity or onboards a large new customer segment, the year-on-year comparison quietly breaks.
A worked example of the shape this produces: a quarter where total attack counts drop by 80 percent while HTTP attacks rise sharply. That is what a booter takedown looks like, or a change in how short low-volume events are aggregated. Hundreds of thousands of trivial stresser-driven bursts disappear from the denominator, the remaining population skews towards deliberate application-layer work, and the two percentages move in opposite directions without the underlying threat shifting at all.
Four checks before you quote any figure:
- The observation window, and whether it matches the window in the report you are comparing it with;
- The customer mix behind the sensors, and whether your sector and architecture resemble it;
- Whether the figures are attempted attacks, detected attacks or mitigated attacks, which are three different numbers;
- Whether the methodology changed since the prior period, which is usually disclosed in a footnote rather than the summary.
The trends that hold across every data set
Despite the disagreements, a handful of trend lines appear in nearly every credible data set, and those are the ones worth building decisions on.
Attacks keep getting shorter. Cloudflare has reported for several consecutive quarters that the majority of the attacks it mitigates end within ten minutes. Most never cross the hour mark. That single trend has more direct buying consequence than any peak-volume record, because it turns time-to-mitigate into the number that decides whether your protection does anything at all.
Peaks keep climbing while the median stays modest. The Spamhaus incident of March 2013, around 300 Gbps, was described at the time as the attack that broke the internet. The 1.35 Tbps memcached reflection attack against GitHub in February 2018, documented by Akamai, reset the ceiling again. Cloudflare reported a 7.3 Tbps attack in May 2025. Meanwhile the typical attack most organisations actually face remains a long way below those figures. The average across a population with that kind of tail is not a planning figure. It is an artefact.
Carpet bombing spreads the load. Rather than saturate one host, the traffic is distributed across an entire prefix, a /24 or wider, keeping per-destination volumes under the thresholds that trigger detection. A defender can watch a per-host graph that looks unremarkable while the upstream link runs out of headroom. Threshold tuning set once at onboarding and never revisited is the quiet failure mode here, and it is worth testing against current attack shapes rather than the ones that were common when the policy was written.
Application-layer and DNS events grow as a share of the total, even in periods when aggregate bandwidth figures flatten. Cheap to launch, hard to distinguish from real users, and invisible on a bandwidth graph.
What the numbers look like by layer
Network and transport
Reflection and amplification still supply most of the bandwidth in the headlines: DNS, NTP, memcached, CLDAP and, more recently, a long tail of exposed UDP services. Booter and stresser services put this capability in reach of anyone with a payment card, which is why counts at the low end swing so violently when law enforcement takes a platform offline. Bandwidth figures dominate coverage because they are easy to understand and easy to compare. They are also the layer most ISPs will blackhole for you, which protects their network and takes your service offline at the same time.
Application layer
A sustained flood of requests against a login endpoint, a site search that runs an expensive query, or a checkout path that touches three back-end systems can exhaust an origin at traffic levels that would not register as an anomaly on a transit graph. Thousands of requests per second, all of them well-formed, many of them from residential proxies. Rate limits tuned to a Black Friday peak will not catch it. This is the category where vendor statistics diverge most, because only providers sitting in the HTTP path can see it.
Authoritative DNS, the layer most often left unmanaged
Authoritative DNS rarely gets its own line in a trend report, which is odd given that an organisation whose name servers stop answering is offline in every practical sense, regardless of how healthy its origin is. Three questions are worth answering this week rather than next quarter: who actually runs your authoritative DNS, what capacity and anycast footprint that provider has, and whether a query flood against your zones would appear on the dashboards your team watches for network traffic. For many organisations the honest answers are “the registrar’s bundled service”, “unknown”and “no”. Query floods and the reflection problem overlap here, which is why getting DNS amplification defences right matters on both sides of the resolver.
Map your own exposure before comparing anyone’s statistics
Write down what you expose at each of the three surfaces: transit links and their committed capacity, public application endpoints and the expensive paths behind them, and authoritative DNS. Then read the reports against that map. A report dominated by HTTP flood statistics is genuinely relevant if your exposure is a public web estate, and close to irrelevant if your risk sits in a transit-connected network with a handful of customer-facing services.
Motive and sector: hacktivism, extortion and collateral damage
Hacktivist activity clusters around political events, and the clustering is predictable enough that the NCSC has issued alerts on pro-Russia hacktivist groups targeting critical national infrastructure. The characteristic pattern is short, loud and well-publicised: a public-sector site knocked over for a few minutes to coincide with a ministerial announcement, a screenshot posted to a Telegram channel, national coverage by lunchtime. Duration statistics and impact statistics part company completely in that scenario. The traffic graph shows a blip. The reputational damage runs for days.
Extortion follows its own rhythm. Ransom demands arrive before an attack, during one, or in place of one entirely, and extortion crews have repeatedly impersonated known threat actor names to borrow credibility. If a ransom note lands, security guidance from the NCSC and from UK law enforcement is consistent on the first point: do not pay. Preserve the note and its headers, report it to Action Fraud and through your own incident process, notify your mitigation provider and your ISP so that thresholds and playbooks are ready, and check whether any claimed demonstration attack actually happened on your telemetry rather than taking the sender’s word for it. Payment funds the next campaign and marks you as a payer; it buys nothing enforceable.
Sector league tables shift with the news cycle. Public sector, finance, healthcare, transport and gaming trade places quarter by quarter, which makes last quarter’s chart a weak predictor of next quarter’s target list. Where your sector sat in the most recent report tells you something about the past three months and very little about the next three.
Turning trend data into protection decisions
Here is where the duration statistic earns its keep. If most events end inside ten minutes, an on-demand model that requires BGP diversion and route propagation may spend the bulk of the attack preparing to act. Always-on mitigation removes that gap and costs more in steady state. That is the real trade-off, expressed in minutes rather than marketing language, and it is the decision the trend data most clearly informs. On-demand still makes sense where the exposure is long volumetric floods against a network with headroom to ride out the first few minutes. It makes much less sense against short bursts timed to a product launch or a political announcement.
Advertised scrubbing capacity deserves the same scepticism. A global aggregate of several terabits per second is the sum of every scrubbing centre a provider operates, which is not the capacity available to traffic arriving at the node nearest you during a regional event. Ask about the node, not the total, and read any headline Tbps figure against your realistic exposure rather than against last year’s record.
ISP-level protection covers more than buyers expect at the network layer and almost nothing at the application layer, and the point where it stops is usually a blackhole route. Knowing where your carrier’s included protection ends is the difference between a layered defence and a gap you discover during an incident.
Build the business case from downtime cost, not attack counts. The structure UK buyers can follow: revenue or contractual exposure per hour (including service credits you would owe), plus staff and recovery cost, multiplied by a realistic annual outage estimate in minutes for your risk profile. Set that against the annual cost of managed mitigation, which in the UK market is typically tiered on committed clean bandwidth and the number of protected prefixes or IPs, with separate charges for on-demand activation events, traffic above commitment, additional protected services and onboarding. The headline monthly price is rarely the figure that appears on the invoice. Ask for a worked example of a bad month before you sign.
Statistics that mislead buyers, and the questions to ask instead
Record peak figures are the statistic most used in sales conversations and the least relevant to most buyers. Unless you are a hosting provider or a carrier, your realistic exposure sits orders of magnitude below the record, and capacity you will never touch is a poor reason to choose one provider over another with better detection and a tighter mitigation SLA.
Percentage growth quoted without a baseline is close to meaningless; a rise from four events to twenty is a 400 percent increase and a rounding error. Year-on-year comparisons across a changed methodology are worse, because they look rigorous. Mitigated-attack counts presented as evidence of market leadership measure sensor coverage, not share.
Live attack maps belong in the same category. They are built from one vendor’s honeypot and customer telemetry, geolocated to source IPs that are frequently spoofed or proxied, and animated for effect. Useful as a conversation starter, not as data, which is covered in more detail in this breakdown of what a live DDoS attack map shows and hides.
Four questions to put to any provider citing its own threat report:
- What customer population produced these figures, and how closely does it resemble my architecture and sector?
- Are these attempted, detected or mitigated attacks, and how do you count a multi-vector campaign?
- What is the usable capacity at the scrubbing centre nearest my traffic, rather than your global aggregate?
- What does the time-to-mitigate SLA start and stop measuring, and what remedy applies when you miss it?
If the answers arrive quickly and specifically, the report is probably honest. If they arrive as another slide of records, you have learned something too.
For a reminder of how badly forecasting in this field can age, and how familiar the recurring themes are, the archived 2007 review from the crystal ball is instructive reading. The NCSC’s denial of service guidance collection remains the most practical free starting point for UK organisations mapping their own exposure.
Frequently Asked Questions
Why do DDoS attack statistics differ so much between vendor reports?
Because each vendor can only report what its own sensors see across its own customers. A CDN’s figures skew towards HTTP request floods, a transit carrier’s towards volumetric traffic, and a scrubbing provider’s towards events its customers diverted to it. Add differing rules on what counts as one attack and the totals stop being comparable.
What is the most important DDoS trend for defenders right now?
Falling attack duration. Cloudflare has reported across multiple quarters that most mitigated attacks end within ten minutes, which pushes time-to-mitigate ahead of raw capacity as the measure that decides whether protection takes effect during the event or after it.
Are DDoS attacks getting shorter or longer?
The aggregate picture is shorter, with a large share of events over in minutes, though long multi-day campaigns still occur against specific targets. Both patterns coexist, so the planning question is which shape your organisation is realistically exposed to rather than which one dominates the chart.
Do live DDoS attack maps show real attack statistics?
They show real telemetry from one provider, visualised. Source locations are derived from IP addresses that are often spoofed or routed through proxies, the sample covers only that vendor’s view, and the animation is designed to be watchable rather than precise. Treat them as illustration, not measurement.
How much scrubbing capacity do record attack figures suggest I need?
Very little, for most buyers. Record figures describe the extreme tail of a global distribution and say nothing about your exposure. Size against your own transit capacity, your application’s concurrency limits and your authoritative DNS provider’s footprint, then ask what capacity is available at the scrubbing location nearest your traffic.
