Dyn and Mirai: The Day the Internet's Phone Book Was Flooded
On Friday, October 21, 2016, the websites themselves were fine. Their servers were up, their code was fine, nobody had broken in. What failed was the lookup: Dyn, the company that ran the managed DNS service — the internet's phone book — for many large sites, was flooded with attack traffic in three waves across one day. One confirmed source of that traffic was Mirai, a botnet built from hijacked cameras, routers and video recorders. Dyn's own two statements and the U.S. Department of Justice record tell a stranger and more useful story than the headlines did — including why the first count of the attack's size was wrong by two orders of magnitude, and why the honest version of who attacked Dyn is narrower than most retellings claim.
The short answer
On Friday, October 21, 2016, Dyn — a company whose Managed DNS service acted as the phone book for many of the internet's best-known websites — was hit by a distributed denial-of-service (DDoS) attack: a flood of junk traffic sent from many places at once to overwhelm a target. The attack came in three waves. The first began at about 7:00 a.m. Eastern Time and was mitigated about two hours later; a second, more global wave began just before noon and was beaten in just over an hour; a third was stopped with no customer impact at all. Dyn never had a system-wide outage — but users routed to its East Coast servers could not reach some customer sites, because the lookup failed even though the sites themselves were healthy. With help from the analysis firms Flashpoint and Akamai, Dyn confirmed that one source of the traffic was devices infected by Mirai, a botnet — an army of hijacked devices controlled as a group without their owners' knowledge — built from internet-connected cameras, routers and video recorders. Five days later Dyn revised its size estimate sharply downward, from tens of millions of IP addresses to up to 100,000 malicious endpoints, and explained why: much of the flood was honest servers retrying. A year later, Mirai's three creators pleaded guilty in federal court to building and operating the botnet. What the record does not say — and this case will not claim — is that those three men personally aimed it at Dyn.
The company: what Dyn did
Dyn ran Managed DNS. DNS — the Domain Name System — is the internet's phone book: when you type a website's name, a DNS server looks up the numeric address (the IP address) your device actually connects to. No lookup, no visit — it does not matter how healthy the website itself is. A managed DNS provider sells that lookup as a service: instead of running their own phone book, many large websites paid Dyn to answer those lookups for them, from Dyn's servers spread around the world.
Those answering points are called POPs (points of presence) — locations where the provider runs its service. Dyn's statement describes East Coast POPs and West Coast users separately, which turns out to matter: this attack hit some answering points hard and left others working, so whether a site "was down" on October 21 depended on where you were asking from.
Dyn described itself, in its statement the day after, as often being "the first responder of the internet" — the company other companies called when they were attacked. On October 21, the first responder was the patient. Why this matters: Dyn was not a website. It was a piece of shared plumbing underneath many websites at once — and shared plumbing is exactly where a single failure stops being single.
The problem: an army of small devices, and a phone book everyone shared
Two separate weaknesses had to exist before October 21 could happen. Neither one, alone, was enough.
The first weakness was the devices. In the summer and fall of 2016, according to the U.S. Department of Justice (DOJ), three young men — Paras Jha, 21, of Fanwood, New Jersey; Josiah White, 20, of Washington, Pennsylvania; and Dalton Norman, 21, of Metairie, Louisiana — created Mirai. Mirai targeted IoT devices — "Internet of Things" devices, meaning everyday objects connected to the internet that are not really computers: in the DOJ's list, wireless cameras, routers, and digital video recorders. The three hunted for both known and previously undisclosed vulnerabilities (flaws in the devices' software) that let them quietly gain administrative, or high-level, access and force the devices into the botnet. At its peak, the DOJ says, Mirai consisted of hundreds of thousands of compromised devices — each one weak on its own, together a flood no single target could shrug off. The trio used the botnet to conduct powerful DDoS attacks, the DOJ says. Then, in the fall of 2016, Jha posted Mirai's source code — the human-readable instructions the program is built from — on a criminal forum. From that point, anyone could build their own Mirai. Hold that fact; it decides what can honestly be said about October 21.
The second weakness was concentration. Every company that outsourced its DNS to Dyn had, without quite deciding to, made the same bet: Dyn stays up, or we are unreachable. Their own servers, their own security, their own backups — none of it was in the chain that failed. An analogy: a street of well-built shops, each with its own locks and alarms, all listed in one phone book. Set fire to the phone book and every shop is "closed" — not because anything happened to the shops, but because nobody can find them.
What this shows: the attack's shape, animated. The flood aimed at the lookup step, not at the websites — which is why so many unrelated sites failed together while their own systems were untouched.
What happened on October 21, 2016
All times are Eastern, from Dyn's statement of October 22, 2016, written by its Chief Strategy Officer, Kyle York. Dyn deliberately gave approximate times and withheld full mitigation detail "to preserve future defenses" — this timeline keeps Dyn's own precision rather than inventing a sharper one.
What this shows: the day's spine, in Dyn's own times. Two damaging waves about five hours apart, a third stopped cold — and the real analysis still to come five days later.
Wave 1 (~7:00 a.m. ET). Dyn's Network Operations Center — the NOC, the team that watches the service around the clock — deals with DDoS attacks routinely, York wrote, but "it quickly became clear that this attack was different." For about two hours, internet users directed to Dyn's East Coast servers were unable to reach some of its customers' sites, including, in York's words, "some of the marquee brands of the internet." Just as important is what did not happen: "Dyn did not experience a system-wide outage at any time — for example, users accessing these sites on the West Coast would have been successful." The phone book had regional editions, and only some were burning.
Wave 2 (just before noon ET). The second wave was "more global in nature" — not limited to the East Coast POPs — and was mitigated in just over an hour, with service restored at approximately 1:00 p.m. ET. Again there was no network-wide outage, though some customers saw extended latency — answers arriving, but slowly — during it.
Wave 3. News reports of a third wave were real, York confirmed — but by then Dyn's defenses were tuned to the attack, and it was "successfully mitigate[d] without customer impact." The pattern is worth noticing: each wave did less damage than the one before, because the defenders were learning the attack's shape faster than the attacker changed it.
What the flood was made of. Dyn's follow-up analysis, published October 26 by Scott Hilton, its executive vice president of product, filled in the mechanics. The attack traffic was "maliciously targeted, masked TCP and UDP traffic over port 53" — port 53 is the numbered channel DNS lookups travel on, and "masked" means the traffic was disguised to look like ordinary lookups, which is what made it hard to filter: during a DNS attack, Hilton noted, it can be difficult to distinguish legitimate traffic from attack traffic. Early observations showed packet-flow bursts 40 to 50 times higher than normal. And one source of it, Dyn confirmed with Flashpoint and Akamai's help, was devices infected by the Mirai botnet.
What this shows: the mechanism behind Dyn's corrected numbers. The attack did not just flood Dyn — it made the internet's own politeness (ask again if you get no answer) flood Dyn too.
The number that shrank. On October 22, Dyn described "a sophisticated, highly distributed attack involving 10s of millions of IP addresses," and said it had observed tens of millions of discrete addresses associated with the Mirai botnet. On October 26, Hilton corrected the record himself: "the estimate at the time of this report is up to 100,000 malicious endpoints." Both statements are Dyn's; the second explains the first. Many attack addresses were spoofed (faked return addresses, so one device can appear as many), and the rest of the illusion was the retry storm: when Dyn's flooded servers could not answer in time, the honest servers asking them — recursive servers, the intermediaries that look addresses up on behalf of users and keep copies in a cache — simply asked again, generating, in Hilton's words, a storm of legitimate retry activity creating 10 to 20 times normal traffic volume across a huge number of addresses. "It appears the malicious attacks were sourced from at least one botnet," he wrote, "with the retry storm providing a false indicator of a significantly larger set of endpoints than we now know it to be." One more number deserves the same honesty: outside reports put the attack's size at 1.2 terabits per second. Dyn's own verdict, in writing, was: "at this time we are unable to verify that claim." This case leaves it exactly there.
The damage
Notice what kind of damage this was. No customer data was stolen from Dyn's customers in this incident, no website was defaced, and no server belonging to a customer was damaged — the DOJ record and Dyn's statements describe a pure availability failure: for hours at a time, across two waves, a large share of the internet's users simply could not reach sites that were themselves running perfectly. The harm was measured in unreachable hours and in trust: every Dyn customer learned, on the same morning, that its availability had a dependency it did not control.
Honesty about scale matters here too. Dyn published no dollar damage figure, and neither did the DOJ releases — so this case prints none, rather than borrowing an estimate from coverage that neither primary supports. What the record does support: the attack was, in York's words, likely "to be seen as an historic attack"; it took a specialist firm's full operations and security teams, plus upstream providers and scrubbing services, most of a day to fully beat; and probing attacks against Dyn continued for days afterward, handled without significant further impact. The deeper damage was demonstrative. October 21 proved, in public, that a botnet of cheap cameras and recorders could bend a core piece of internet plumbing — and Mirai's source code was already public, so the proof came with instructions.
How they recovered
During the day itself, recovery was mitigation, wave by wave. Hilton's October 26 summary lists the techniques Dyn used, alongside help from its upstream providers, who blocked some of the attack flow before it reached Dyn: traffic shaping (deliberately slowing and prioritising incoming traffic so the important kind gets through), rebalancing traffic by manipulating anycast policies (anycast is the technique of announcing the same service address from many locations so each user is answered by a nearby one; changing the policies shifts load between those locations), internal filtering (dropping traffic matching the attack's fingerprint inside Dyn's own network), and deployment of scrubbing services (specialist cleaning stations that wash attack traffic out of a stream and pass the clean traffic on). The results are in the timeline: wave 1 took about two hours to mitigate, wave 2 just over an hour, wave 3 nothing at all.
After the day, Dyn said it was working with other infrastructure providers on mitigation strategies so future attacks could be brought under control faster, and it cooperated with an ongoing criminal investigation — which is why Hilton, on the record, would "not speculate regarding the motivation or the identity of the attackers." That restraint is part of why this case is written the way it is.
The legal ending came a year later. On December 8, 2017, Jha, White and Norman each pleaded guilty in the District of Alaska to conspiracy to violate the Computer Fraud and Abuse Act in operating the Mirai botnet, the DOJ announced on December 13, 2017. On September 18, 2018, all three were sentenced in Alaska to five years' probation, 2,500 hours of community service, and $127,000 in restitution, and they voluntarily abandoned significant amounts of cryptocurrency seized during the investigation. Jha separately pleaded guilty in New Jersey over his own 2014–2016 DDoS attacks on Rutgers University, which had repeatedly shut down the university's central authentication server; for that he was ordered to pay $8.6 million in restitution and serve six months of home incarceration, five years of supervised release and 2,500 hours of community service. Read the attribution one more time, carefully, because it is the point: the convictions were for creating and operating Mirai and for the Rutgers attacks — the primaries do not charge the trio with the Dyn attack itself, and Dyn said only that Mirai-infected devices were one source of its traffic, in a season when Jha had already given the source code away.
The fix: technology and controls that came out of it
Do not let one provider be the only way to find you. The structural lesson of October 21 is written in Dyn's own geography: the sites stayed reachable from the West Coast because that part of Dyn still answered. Redundancy works when it is genuinely independent. For DNS, that means a second, independent way to answer lookups — a secondary DNS provider on separate infrastructure, or an in-house fallback — arranged before the bad day, because you cannot sign up for a new phone book while the old one is burning and nobody can find your site to begin with. A dependency you share with half the internet is still a single point of failure; it is just a famous one.
Design for the retry storm, not just the attack. Dyn's most transferable measurement is the 10-to-20-times figure: when a lookup service degrades, its own honest clients multiply the load. Capacity plans that only model attack traffic miss the bigger half of a DNS incident. The controls follow from the mechanism — longer-lived cached answers where the service allows it (a resolver holding a still-fresh cached answer does not need to ask again), rate-limiting that treats repeated identical questions gently instead of as fresh demand, and anycast spread wide enough that no single region's failure becomes the story. Dyn's wave-by-wave improvement also shows the value of rehearsed mitigation: shaping, filtering, anycast rebalancing and scrubbing were playbook moves, executed faster each time.
Fix the device problem at the device. Mirai's raw material was hundreds of thousands of cameras, routers and recorders whose owners never knew they were conscripted. The DOJ's plea announcement itself pointed readers to its published cybersecurity best practices for IoT devices — a quiet admission that the botnet's supply chain is a hygiene problem: devices must not ship or sit on networks with guessable or unchangeable administrative access, known vulnerabilities need patching, and equipment that can never be updated does not belong where a botnet can recruit it. Every unpatched camera is, after October 21, best understood as somebody else's weapon in storage.
Count honestly, in public, fast. The last control is not technical. Dyn's October 22 statement gave a number its own analysis replaced five days later — and Dyn published the correction itself, with the mechanism (spoofing plus retries) explained, and refused to endorse the 1.2 Tbps figure it could not verify. That sequence is why the record of this incident is still usable a decade later: the first statement said what was known, labelled what was not, and the follow-up narrowed it. Incident communication that overclaims early and quietly forgets it later teaches customers to discount everything you say in the next incident.
What this means for you
Enterprise network engineers
Your version of Dyn is every shared service you do not run: the DNS provider, the identity provider, the cloud region's control plane, the one SaaS status page everyone needs during an outage. October 21's test is a single question — if this provider stops answering for three hours, what exactly stops working for us, including the tools we would use to fix it? Write the answer down before someone else runs the experiment. For DNS specifically: know who answers your lookups, know whether a second independent path exists, and know how long your cached answers keep you findable while the provider is down. And take the retry storm home with you — your own applications retry too, and an internal service that degrades will be finished off by its clients' retries unless back-off (waiting longer between attempts) is built in.
Hyperscaler engineers
You are Dyn in this story: the shared plumbing. The case hands you three disciplines. First, blast-radius honesty — Dyn could truthfully say "no system-wide outage" because its regional structure genuinely contained wave 1; design so that sentence is true for you too, and be able to prove it per region, per POP. Second, measurement under fire: Dyn's packet counters saw 40–50x bursts excluding traffic already blocked upstream, and it said so — know which of your numbers are floors, not totals, before an incident makes you quote them. Third, the correction culture: a revised estimate published in five days, with the false-indicator mechanism named, is what turned a marketing embarrassment (the first number was wrong) into the incident's most cited technical contribution. The retry-storm finding has shaped DNS resilience thinking ever since — because Dyn published it instead of burying it.
Try it yourself
Dyn's customers discovered that their availability was not their own number — it was their number multiplied by Dyn's. You can feel that maths with Packet Path's free SLA and Availability Calculator: enter 99.9% for your own service, then chain it behind a second 99.9% service it depends on, and watch the combined figure — and the monthly downtime budget — fall. It is the whole case in one calculation: a dependency you do not control is still your outage.
Sources
Primary source — Dyn, "Dyn Statement on 10/21/2016 DDoS Attack," Kyle York, Chief Strategy Officer (October 22, 2016). All wave times, the East Coast impact, the no-system-wide-outage finding, the tens-of-millions first estimate, and the Mirai-as-one-source confirmation (with Flashpoint and Akamai) are from this statement. The original hub.dyn.com post now redirects following Dyn's absorption into Oracle; the full text was read for this case in the contemporaneous seclists.org archive of the statement: https://seclists.org/interesting-people/2016/Oct/59
Dyn follow-up analysis — Scott Hilton, Executive Vice President of Product (October 26, 2016): the up-to-100,000-endpoints revised estimate, masked TCP and UDP traffic over port 53, 40–50x packet bursts, the 10–20x legitimate retry storm, the mitigation list (traffic shaping, anycast policy rebalancing, internal filtering, scrubbing services, upstream blocking), and the statement that Dyn was unable to verify reports of 1.2 Tbps — read for this case in CSO Online's October 27, 2016 report quoting Hilton's post directly: https://www.csoonline.com/article/558603/ddos-attack-against-overwhelmed-despite-mitigation-efforts.html
U.S. Department of Justice, District of New Jersey, press release, "Justice Department Announces Charges And Guilty Pleas In Three Computer Crime Cases Involving Significant Cyber Attacks" (December 13, 2017) — Mirai's creation in summer–fall 2016, its IoT targets and peak size, the December 8, 2017 guilty pleas, the fall-2016 source-code posting, and Jha's Rutgers guilty plea: https://www.justice.gov/usao-nj/pr/justice-department-announces-charges-and-guilty-pleas-three-computer-crime-cases
U.S. Department of Justice, District of New Jersey, press release, "Computer Hacker Who Launched Attacks On Rutgers University Ordered To Pay $8.6m Restitution; Sentenced To Six Months Home Incarceration" — the September 18, 2018 Alaska sentencing of all three defendants (five years' probation, 2,500 hours of community service, $127,000 restitution, abandoned cryptocurrency) and Jha's New Jersey sentence ($8.6 million restitution, six months' home incarceration): https://www.justice.gov/usao-nj/pr/computer-hacker-who-launched-attacks-rutgers-university-ordered-pay-86m-restitution
Attribution note: no primary source above states that Jha, White and Norman personally launched the October 21, 2016 attack on Dyn, and no affected customer is named in Dyn's statement — this case claims neither.
- A shared dependency is a shared fate: Dyn's customers' own servers were never touched — when the DNS lookup failed, their healthy sites were unreachable anyway.
- Dyn's first size estimate was wrong in an instructive way: 'tens of millions of IP addresses' on October 22 became 'up to 100,000 malicious endpoints' on October 26, because spoofed addresses and a storm of legitimate retries made the crowd look far bigger than it was.
- The retry storm is part of the attack whether the attacker plans it or not: when DNS answers stop, honest servers ask again — Dyn measured 10 to 20 times normal traffic volume from legitimate retries alone.
- Attribution discipline matters: Dyn confirmed Mirai-infected devices were one source of the traffic, and the DOJ convicted Mirai's three creators for building and operating the botnet — neither primary source says those three men personally launched the Dyn attack, and this case does not claim it.