Zenpanik
The Red Queen's race: how Apple, Google and evolution teach systems to stop attacks earlyBy Pavel Nikolov, November 9, 2026

Part 4: Locks at machine speed

Part 4 of 6 in a deep dive on machine learning that predicts attacks and stops them early, with Apple as the main case. Start with Part 1 if you are new to the series.

Part 2 showed how Apple locks the device, and Part 3 how defenders predict which bug will be used against it. Between the device and the attacker sits a third layer: the locks on every door. In 2026 most of those locks are models. They decide in seconds whether a sign-in, a request, a message or a file gets through, and attackers now use AI to test them faster than any person could. This part is about those locks, how they work at machine speed, and what happens when they become the target.

The control room moved to machine speed

Microsoft's Digital Defense Report, published in October 2026 and drawing on more than 165 trillion security signals a day, puts the shift in one line: for sophisticated attackers, AI reduces the attack chain from days to seconds (Microsoft, Security MEA). Its conclusion for defenders is that attacks at machine speed can only be met by defenders moving at the same speed, with people working alongside agentic AI systems.

The other 2026 reports agree. Verizon found that exploiting a vulnerability overtook stolen credentials as the top way in for the first time in 19 years, at 31% of breaches (Verizon). IBM put the average breach at $4.99 million and found that security AI and automation saved $1.93 million per breach (Help Net Security, Forkast). In June, Google said its triage agent had investigated more than 5 million alerts, cutting a typical 30-minute manual analysis to about 60 seconds (TechInformed).

MinutesManual triage of one alert30 minAverage attacker breakout, 202529 minGoogle's triage agent, 2026about 1 minManual triage time and agent time from Google; breakout time from CrowdStrike.
Figure 9. The race in one chart. By the time a person finishes triaging one alert, the average attacker has already moved to a second system. Sources: TechInformed on Google, CrowdStrike.

Speed is half of it. Microsoft calls the other half the primary strategic variable: how well a defender correlates signals across its own systems and with others (Microsoft). That idea runs through the rest of this series, from the sign-in page to the vault door.

Four digital locks

The network edge. Floods are already answered without people. In September 2025 Cloudflare blocked a 22.2 terabit-per-second flood that lasted 40 seconds, detected and mitigated autonomously (SecurityWeek), and it mitigated 23.2 million network-layer attacks in the first half of 2026 (BleepingComputer). Exposed cloud workloads were attacked 5.3 hours after they appeared, on average (Microsoft).

Sign-ins. In Microsoft's 2026 data, user execution, valid accounts, social engineering and phishing made up 73.3% of initial access attempts, and more than half of intrusions led to further credential theft. Its incident responders saw the same four steps again and again (Microsoft).

Run 1: the chain runs to the endRun 2: a model flags the second step1. Human identityphishing, vishingor a stolen password2. Non-human keysservice principals,API keys, tokens3. Escalationmore privilegeswith existing access4. Objectivesdata theft orcommand and controldata goneunusual use of a service key: blockedThe chain is Microsoft's; the point where the model acts is an example.
Figure 10. How identity attacks unfold, according to Microsoft Incident Response in 2026. The earlier a model breaks the chain, the less there is to clean up. Source: Microsoft Digital Defense Report 2026.

Passkeys and phishing-resistant MFA lock most of this door. The model's job is the second step: a service key that suddenly signs in from a new place, or an API token used for something it has never done, is an anomaly a classifier can score in seconds and a person reading logs finds days later.

The help desk. When the sign-in page holds, attackers call the person who can reset it. Microsoft found that 93% of voice phishing attacks kept the victim on the line long enough to begin the social engineering (Microsoft), and in Gartner's 2026 survey of CISOs, 41% reported a deepfake used on an audio call (Keepnet on Gartner). A model cannot reliably hear a cloned voice yet, so the lock here is process: call back on a number already on file, never reset MFA on one call, and show the agent the identity model's risk score first.

Pages and messages. Microsoft detected more than 145 million QR-code phishing attacks over the year (WorkOS). Google's Gemini Nano reads suspicious pages inside Chrome and gives Safe Browsing a verdict on the computer itself (Infosecurity Magazine), the same on-device pattern Apple uses for its spam filter and Call Screening.

Keeping the locks honest

Every lock a model controls has a blind spot. When it blocks a sign-in or drops a request, nobody learns whether that was really an attack, so the next model trains only on what the last one let through. Stripe hit this early with its fraud models. Michael Manapat, who led machine learning for Stripe Radar, described the fix: let a small random share of would-be-blocked events, 5% in his example, get a real outcome, and count each one 20 times in training so the data looks like a world where nothing was blocked (Manapat's slides).

Sign-insAllowedBlocked: no outcome ever arrivesSampled at random for review: 5% in Stripe's exampleModelscores eacheventOutcomesconfirmed attackor false alarmsampled eventscount 20 timesRetrain on the outcomes
Figure 11. Keeping the lock honest: the model blocks most risky events, but a small random sample still gets a real outcome, so the next model learns about the attacks the last one stopped. Adapted from Michael Manapat's description of Stripe Radar; the 5% rate and weight of 20 are from his example.

In security, letting an attack through on purpose is rarely acceptable, so the practical version is review rather than release: send a random sample of blocked events to an analyst, sample borderline scores more often than obvious ones, and weight those labels the same way. The same rule returns in Part 5, where the blocked events are dismissed camera alarms.

When the lock is the target

In July 2026, Sysdig documented JADEPUFFER, which it describes as the first ransomware extortion run end to end by an AI agent. The agent got in through an unauthenticated code execution bug in Langflow, an open-source tool for building AI workflows, then found its target, harvested credentials, set up persistence and delivered the ransom demand (GCN). Microsoft tracks the group as Storm-3168 and saw it use compromised service principals, step two of the chain above, to delete more than 100 Azure storage accounts (The Hacker News, Windows Report). The way in was an AI tool, and the speed came from not waiting for a person.

IBM's 2026 report measured how fast AI systems became part of the attack surface (Help Net Security, SureCloud).

Breaches involving an AI model or app2025: 13%2026: 21%Incidents involving shadow AI2025: 20%2026: 43%AI breaches without proper AI access controls2026: 92%
Figure 12. AI systems became part of the attack surface in a single year. Source: IBM Cost of a Data Breach 2026, via Help Net Security and SureCloud.

Evasion changed too. Classic malware hides its code from the scanner. Google described families such as HONESTCUE that ask Gemini for code while they run, so the malicious logic is never in the file a scanner inspects (SiliconANGLE).

programon diskscanner: nothing foundasks for codeAI model APIwrites the payloadgenerated codememorythe code runs hereSchematic of the pattern Google described for HONESTCUE.
Figure 13. Why file scanning misses the 2026 malware. The program on disk is clean; the payload exists only after the model writes it. Based on Google Threat Intelligence Group, via SiliconANGLE.

What still works is watching behavior rather than files: a process with no reason to call a model API that does, or code that appears in memory without touching disk. Part 5 meets the same trick at the physical door, where a forged identity, not a crowbar, opens the loading dock.

When the lock jams

Attackers are not the only risk. On November 18, 2025, Cloudflare's Bot Management system, which uses a model to score every request for whether it comes from a bot, took down much of Cloudflare's network. A database permissions change made the query that builds the model's feature file return duplicate rows. The file, refreshed every few minutes and pushed to every machine, doubled in size, passed a hard limit in the proxy software and crashed it (Cloudflare). It normally used about 60 features against a limit of 200 (EveryWAN), and most traffic recovered only at 14:30 UTC, more than three hours later (CyberInsider).

All devices at onceCanary group firstEvery device takes the bad update.Three devices fail. The rollout stops.Schematic. Each square is a device; the canary share is illustrative.
Healthy deviceDevice hit by a faulty update
Figure 14. The same faulty update or model file, pushed two ways. A staged rollout turns a global outage into a handful of broken test machines. Schematic, based on Cloudflare's and CrowdStrike's post-incident changes.

Microsoft's 2026 report states the principle: resilience by design assumes that failures will occur (Microsoft). Ship every new model, rule or feature file to a canary group first, and validate generated files as strictly as input from outside.

From software locks to steel ones

Every lock in this part is software. The oldest locks are not. Doors, vaults, safe deposit boxes and the trucks that carry gold and servers are guarded by cameras, sensors and people, and in 2026 those are models too. The attackers are the same, and so are their tools: forged identities, AI-written emails and agents that do the research. Part 5 follows them to the physical door.

Continue to Part 5: The physical door