Global Internet Outage: How to Tell an ISP Failure From a Cyberattack

2026-11-18来源 dev.to DevOpsEnglish
Global Internet Outage: How to Tell an ISP Failure From a Cyberattack

"The internet is down" covers six different problems, and only one of them deserves the word global. Deciding which one you're in is the first job of an incident, because the fix, the people you call, and the customer message all differ by class.

  • Local ISP or customer-premises failure. Your office fiber, a modem, an ONT, a regional ISP footprint. Everything is unreachable from one location while the same services work from a phone on cellular. Most common class by a wide margin.
  • Regional carrier or backbone disruption. A transit provider, subsea cable or metro ring loses capacity. Multiple independent networks in one geography lose the same paths.
  • DNS or BGP incident. A bad route announcement or a resolver failure makes healthy services unreachable for millions of users with no intrusion anywhere.
  • Cloud or SaaS dependency outage. One provider's region, AZ or control plane fails and every product built on it degrades at once.
  • DDoS or confirmed cyberattack. Resource exhaustion by traffic. An availability attack, which is a different thing from an intrusion that compromises systems or data.
  • Coincident reports that only look global. Several unrelated incidents in the same hour, user-report maps light up, social media supplies the word "global." That last class is underrated and it's usually the answer. Thousands of providers run incidents every day; on any given afternoon several overlap. A check across nine provider status pages on 17 November turned up exactly that shape — Cloudflare with R2 errors from South America and some AI models, Apple with iCloud sign-in problems, Twilio with eight carrier-specific incidents across seven countries, OpenAI with a delayed Costs API. Microsoft, GitHub and Google Cloud reported nothing; AWS showed two long-running regional issues in UAE and Bahrain.

Every one bounded by product, region or carrier. None attributed to malicious activity. A genuine global outage needs a shared layer to fail: a backbone, a major DNS or BGP path, or a cloud control plane the others depend on.

The diagnosis matrix

Work it top to bottom. Starting at "have we been hacked" and working backwards burns the first hour chasing the rarest explanation.

Symptom Likely layer First check (<5 min) What confirms it What isn't enough
Unreachable from one office; cellular works Local ISP / CPE Modem or ONT lights; ISP status page; a second site on the same ISP ISP notice naming your footprint A spike on a national outage map
Several sites in one region lose the same destinations Regional carrier Traceroute from two providers in the region Multiple independent networks losing the same paths; route telemetry A list of unrelated SaaS apps being slow
Sites "won't load" but IPs ping; SERVFAIL or NXDOMAIN DNS Query a second resolver; test authoritative servers for your domains Resolver or authoritative failures plus a provider acknowledgement A user saying "the website is down"
Specific destinations unreachable from many networks; paths change abruptly BGP / routing Looking-glass or route-collector check for your prefixes and theirs Abnormal announcements or withdrawals plus an operator statement Social posts speculating about a hack
One vendor's products degrade together; your infra is healthy Cloud / SaaS Provider status page and your tenant's service-health view Provider incident naming region, service, impact Downdetector alone
Edge floods with traffic; latency climbs DDoS Edge and WAF dashboards; provider DDoS notice Traffic evidence plus provider or security confirmation and a mitigation record Outage size or suddenness
Unexpected admin activity or config changes alongside the outage Intrusion Identity logs, change audit, EDR; open a security incident Forensic evidence of compromised accounts, or a provider disclosure Any loss of availability by itself

Outage and security incident are different words

An outage is loss of availability. A security incident is evidence that confidentiality, integrity or control was compromised.

Outage Security incident
What failed Reachability or capacity Trust: an account, system, dataset or control plane
Who declares it Provider status page, NOC, your SRE lead Your security team, a provider disclosure, or a regulator
Evidence Status incident, route telemetry, resolver errors Forensic artifacts, compromised credentials, attacker activity in logs
What ends it Service restored, confirmed by synthetic checks Containment, eradication, recovery, written root cause
Customer message "Provider X is degraded; here's the workaround and next update" Legal and compliance review before any statement

A DDoS sits in both columns — an attack by intent, an outage by effect. And a vulnerability is a weakness, not an incident; it becomes one when someone exploits it. Treat an unpatched edge device as a reason to investigate during an outage, not as a reason to announce an attack.

The rule: use the word "attack" only after a responsible primary source confirms it. Everything before that is an outage of unknown cause, and that phrase is fine to put in a customer notice.

First 15 minutes, first hour, first four hours

0–15 min — is this us, our ISP, or a provider? Confirm the symptom from a second network. Check your ISP's status page and your own synthetic monitors. Identify one known-good path for the most critical flow. Open a channel and name an incident lead. The output is a scope statement with the evidence behind it, not a cause.

15–60 min — which layer, and what do we tell customers? Run the matrix. Pull the dependency map and list every customer-facing flow touching the affected layer. Publish one message with known scope, workaround, and next update time. Route everything else through the incident lead.

1–4 hours — which fallback, and what do we preserve? Activate only fallbacks tested in the last quarter. An untested failover at hour two turns one outage into two. Preserve logs, route snapshots, DNS query logs and edge captures with timestamps, because if the classification later moves toward attack, the logs that would settle it have usually rotated away.

Day 1–2. Rebuild the timeline from evidence rather than memory. Review which flows failed, which fallbacks worked, which didn't exist.

How much redundancy is worth buying

Enough for the flows whose failure costs more than the redundancy. This runs against the commercial interest of anyone selling resilience work, so it's worth saying plainly.

Where independence pays: a second ISP on a different physical path at a site that can't run on cellular; secondary DNS at a different provider with the zone actually replicated and tested; a provider-independent path for the one or two flows whose outage starts a regulatory clock or costs measurable revenue within the hour.

Where it costs more than it saves: two clouds for a workload that tolerates four hours down; dual-vendor SaaS for email or chat; any failover you can't test quarterly. Every added provider is also added surface — a second identity integration, a second credential set, a second patch cycle.

Five decisions cover most of it: map dependencies before buying anything; define a known-good path per critical flow and who may switch to it; delete untested failovers from the runbook until they're exercised; retain evidence long enough to classify an incident days later; and decide in advance which status sources your incident lead trusts, so nobody classifies an incident from a social feed.

A team that gets those right has most of the resilience it will ever use. A team that buys a second cloud before doing the dependency map has bought a second set of problems.

Full write-up — the November 17 provider-by-provider check, the full matrix, and the industry dependency chains: teamvoy.com/blog/global-internet-outage

Written by the Teamvoy team. More engineering writing at teamvoy.com/blog.

原文:https://www.a8q9.cn/en/article/global-internet-outage-how-to-tell-an-isp-failure-from-a-cyberattack-365cb4.html

相关文章