A hosting provider can lose a customer's connection to a five-minute flood, recover before an engineer opens the alert and face another flood an hour later. Neither incident has to reach a terabit per second. A saturated access link, a packet-processing bottleneck or exhausted connection state can turn an unremarkable bandwidth figure into a real outage.

That is the operational reading of Cloudflare's third-quarter 2026 threat report. The data describes more network-layer attacks, most of them short and modest in bandwidth. It does not mean large attacks have gone away.

Millions of attacks, minutes of disruption

From July to September 2026, Cloudflare says it autonomously mitigated 18.2 million network-layer DDoS attacks, up 208% from the equivalent quarter of 2025. Ninety-six percent ended in under ten minutes and 98% stayed below 1 Gbps. The provider separately reports 14.4 trillion mitigated HTTP DDoS requests; those are requests, not 14.4 trillion distinct incidents.

At the other extreme, network-layer attacks exceeding 1 Tbps fell from 805 in the second quarter to 251 in the third. The decline coincided with law-enforcement operations against major botnets, but coincidence cannot quantify their causal contribution. These are observations on Cloudflare's protected network, not a census of attacks worldwide.

Duration changes the mitigation problem. If an attack is over before a human responder has confirmed the alert and requested upstream filtering, accurate post-incident analysis does little to preserve availability during the event.

Why one gigabit is an arbitrary threshold for a customer

Imagine a service behind a 100 Mbps circuit. A 400 Mbps flood offers four times that link's capacity even though it falls well below the 1 Gbps level used in the report's distribution. A firewall on the destination server cannot restore bandwidth already consumed before packets arrive.

Packet rate introduces a separate constraint. Small-packet traffic can stress queues, CPU cycles and early receive processing without consuming impressive amounts of bandwidth. Connection-oriented applications also have state tables, admission limits and worker pools that can be exhausted independently of line rate.

The difference is well established in the IETF's analysis of denial-of-service mechanisms, RFC 4732. The novel observation in the recent figures is the mix of frequency and duration, not a new definition of how networks fail.

A statement such as 'protection up to X Tbps' is consequently incomplete. Customers also need to know where filtering happens, which traffic characteristics can be recognized and how much service remains usable during an incident.

UDP attacks are changing shape

Cloudflare reports a 215% quarter-over-quarter increase in DNS amplification attacks and a 122% rise in general UDP floods. CLDAP-related activity fell steeply over the same period. Those statistics concern attack categories observed by the provider; they do not imply that ordinary UDP traffic has become malicious.

Reflection and amplification attacks exploit an asymmetry: a small request with a spoofed source address prompts a larger response to the victim. Legitimate-looking DNS responders may send traffic the destination never requested.

For a game server or real-time communication application, a blanket UDP block can remove both the attack and the service. More selective defenses examine expected destination behavior, protocol structure and the relationship between incoming packets and legitimate sessions. A generic block rule is not equivalent to a protocol-aware mitigation policy.

The response-time problem

A manual incident process includes several delays: telemetry reaches a monitoring system, a threshold triggers, an operator investigates and a change finally propagates. An attack lasting three minutes may be nearly over by the time those steps finish.

Automation can close part of that gap, but an anomalous traffic spike is not proof of hostile intent. A popular game update or a successful product launch can create abrupt, legitimate demand. Poorly scoped automation may reject more real users than the attack itself would have displaced.

The defensible objective is narrow intervention with measured effects. Identify what differs between hostile and legitimate flows, enforce the smallest useful control, watch accepted traffic and remove the rule when supporting evidence fades. A rising drop counter proves packets were discarded; it does not prove customers remained connected.

The UK's National Cyber Security Centre guidance on preparing for denial-of-service attacks emphasizes preparedness and coordination. That advice has particular force when the relevant incident may finish before the first conference call.

Put mitigation before the bottleneck

A typical hosting path passes through transit, an upstream filtering service, a routing edge, a host dataplane and an application. Each layer can reject traffic, but not every layer can protect all the preceding infrastructure.

If the uplink is congested, host-level filtering may preserve server CPU while customers still experience packet loss. If bandwidth remains available but small packets exhaust receive processing, early and efficient host or edge filtering can make a decisive difference. If requests are network-valid but consume expensive application work, a bandwidth-based threshold may miss the failure entirely.

This is also why an outage should not be classified as DDoS simply because customers cannot connect. DNS, routing and CDN incidents have similar symptoms but different remedies. Our analysis of DNS, BGP and CDN failure modes covers that distinction.

Measure disruption, not just attacks detected

A useful DDoS scorecard tracks time to effective mitigation, packet loss or application errors experienced by legitimate clients and evidence of false positives. It also records where the bottleneck occurred. A network team can block millions of packets while an application remains unavailable for unrelated reasons.

Attack count is not business impact. A hundred incidents intercepted before the access link is affected may matter less to customers than one five-minute interruption during a major event.

Cloudflare's figures do not establish the attack distribution of every hosting network. They do show why the maximum attack a provider can absorb is a weak proxy for everyday protection. The practical test is whether filtering reacts before a short flood disrupts the customer, and whether legitimate traffic survives that reaction.