Zenpanik
Deep Dive

The Red Queen's race: how Apple, Google and evolution teach systems to stop attacks early

Attackers now move faster than people can patch or respond. The defenders who keep up predict the next attack and act on their own, in seconds. Apple, Google, Microsoft and Cloudflare show how, from the chip to the network.

By Pavel Nikolov, November 9, 2026
flaggedRisk score of the groupblock
Normal traffic wanders at random. An attack moves together, and its risk score climbs until it crosses the line and the group is blocked.

In 2025, the average exploited vulnerability was attacked seven days before its patch came out. Mandiant, drawing on more than 500,000 hours of breach investigations that year, puts the mean time to exploit at minus seven days, down from 63 days in 2018 and 2019 (Google Cloud). In December 2025 a botnet hit Cloudflare's network with 31.4 terabits per second of traffic, the largest DDoS attack on record (TechRepublic). Cloudflare's earlier record, a 22.2 Tbps flood that lasted 40 seconds, was blocked by its systems with no person involved (SecurityWeek).

Attackers have started to automate too. In September 2025, Anthropic caught a group it assesses as Chinese state-sponsored using its Claude Code agent to run an espionage campaign against about 30 organizations, with the AI doing 80 to 90% of the tactical work (Anthropic).

Attacks now move faster than a person can read an alert. This deep dive is about the defenses that keep up: machine learning that predicts which bug will be exploited, scores a login, a file or a page before it can do harm, and acts on its own when there are only seconds. Apple is the main case, because it combines prediction, randomness and on-device models from the chip to its services. Google, Microsoft, Cloudflare and Stripe fill in the rest.

Why it is a race

In 1973 the evolutionary biologist Leigh Van Valen published a pattern from the fossil record. A group of related species is about as likely to die out after ten million years as after one, so age buys no safety. His explanation, the Red Queen hypothesis, is that predators, prey and parasites keep evolving against each other, and every species has to keep changing to hold its place (University of Chicago). The name comes from the Red Queen in Through the Looking-Glass, who has to run as fast as she can to stay where she is.

Security follows the same dynamic. Google blocked 1.75 million policy-violating apps from Google Play in 2025 and made publishing harder with developer verification and pre-review checks. Over the same period, the malicious apps its Play Protect scanner found outside the store rose from 5 million in 2023 to 27 million in 2025 (Reuters, via Business Ghana). The attackers did not stop. They moved.

A rule set falls behind in this race, because each rule describes an attack that has already happened. A model trained on recent data can catch the next variant if it resembles the last one, and a model retrained every week keeps up with the variant after that.

Security is a system of doors

Think of a company's security as a building with many doors. There is the network edge, the sign-in page, email and the web, the help desk phone, the devices people carry, the software the company pulls in, the AI agents it runs and, oldest of all, the physical doors, vaults and loading docks. An attacker needs one door that opens. The defender has to keep every door locked, all the time, while attackers keep trying new keys.

Your systemsdataappspeopleNetwork edgefloods filteredEmail and webscams judged on deviceIdentitysign-ins scoredHelp deska person on the phonebreachEndpointsmemory tags, LockdownSupply chainbugs fixed earlyAI agentsmisuse cut offPhysicalsensors fused
An attack trying a doorA door a model watches and locks
Figure 1. Security as a system of doors. Machine learning keeps seven of them locked at machine speed. The help desk is the door a cloned voice talks its way through: in 2025, voice phishing was the most common way into the cloud environments Mandiant investigated (Google Cloud community).

Machine learning helps in two ways. It watches each door and notices a key that does not fit, such as a sign-in from an impossible place or a page that behaves like a scam. And it predicts which door attackers will try next, so the defender can reinforce it before the attempt. The rest of this series goes door by door.

Run 1: reactive controls onlyRun 2: a defense acts at the doorPredictive controls act hereReactive controls act hereStagingExploit or botnetbeing preparedGooglezero-day fixed firstAccessLogin, sign-upor phishing linkMicrosoftMFA stops 99%First attemptExploit fires,flood startsCloudflareDDoS auto-blockedLossData stolen orservice downBreachor outageClean-upIncident response,patch, restore14-day mediandwell timelossstoppedThe amber dot is one attack. Each example below the arrow is covered later in this series.
Figure 2. Where defenses act along an attack. In run 1 the attack reaches the loss and the team cleans up afterwards; in 2025 the median intruder went unnoticed for 14 days (Mandiant). In run 2 a defense stops the attack at access, and the later stages never happen.

Three moves the best defenders make

Predict before it happens
Score the vulnerability, the login, the file or the page before it is used. EPSS, a machine learning model maintained by FIRST, estimates the chance that each published vulnerability will be exploited in the next 30 days (Empirical Security). Google's threat intelligence team saw attackers staging an exploit before anyone knew which bug it used, and its Big Sleep agent found the bug in time (The Hacker News).
Make the target random
Attackers need patterns as much as models do. Apple tags every memory allocation with a random secret and re-seeds the generator that picks the tags often, so an exploit cannot guess the next tag (Apple Security Research).
Learn from the whole network
One company sees a few attacks. Microsoft processes more than 100 trillion security signals a day (Microsoft), Cloudflare mitigated 23.2 million network-layer DDoS attacks in the first half of 2026 (BleepingComputer), and Apple watches for the same spyware across all of its devices. A pattern that is rare for one company is common across a network, so the model sees it sooner.

Who guards which door

DoorHow attackers try itWho guards it wellPublished result
Network edgeDDoS floods; exploits against VPNs and firewallsCloudflare autonomous mitigation23.2 million network-layer attacks mitigated in H1 2026
IdentityPassword spraying, stolen credentials, token theftMicrosoft38 million identity risk detections analyzed a day
Email and webPhishing and scam pages written by AIGoogle Chrome with Gemini Nano; iOS 26 filtersGoogle catches 20 times more scam pages in Search
Help desk and phoneCloned voices asking for a password or MFA resetPeople and process, helped by modelsVoice phishing was the top way into cloud environments, at 23%
Endpoints and devicesZero-click exploit chains, mercenary spywareApple MIE, Lockdown Mode and threat notificationsAlerts sent to users in over 150 countries since 2021
Software and supply chainUnpatched bugs, poisoned packages, malicious appsEPSS; Google Big Sleep; Project GlasswingEPSS needs about 8 times less patching than CVSS 7+ for similar coverage
AI agents and toolsAgents told to attack; backdoored MCP toolsAnthropic and Google threat intelligenceGTG-1002 disrupted after a handful of intrusions
PhysicalForced entry, fake pickups, drones, talk at the loading dockFused cameras, radar and access controlCargo theft losses up 39% in 2026 while incidents fell

The series then follows an attack inward, with AI on both sides at every step. Part 2 covers the device: Apple's random memory tags, Lockdown Mode, models that run on the phone, and the bounties Apple pays people to attack it. Part 3 covers the exploit, and 2026, the year attackers got AI, from EPSS to Google's Big Sleep and Project Glasswing. Part 4 covers the locks: the models that guard sign-ins, the network edge and the help desk at machine speed, and what happens when attackers target those models or they fail on their own. Part 5 ends at the oldest door, the physical one, where cameras, radar, Wi-Fi and badge readers can be combined into a single physical fingerprint, and where gold, vaults and truckloads of chips are now attacked with the same AI tools. It compares the companies that sell these locks and what six markets spend on them. Part 6 closes with an evolutionary playbook that uses my open-source library genetic_xgb, and the six findings from 2026 that held up across sources.

Continue to Part 2: Apple, the device