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

Part 6: The playbook

Part 6 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.

Five parts followed one attack inward, from the device to the exploit, the locks and the physical door. Attackers adapt, locks can be fooled, and locks can jam on their own. This last part asks what a defender should build in response, and borrows the answer from the oldest defense system there is.

What evolution teaches

Nature does not answer adaptive attackers with one perfect lock. The immune system keeps a varied population of detectors and keeps replacing them, and Taleb's 1/N rule says the same in statistical terms: keep each error small and separate, so no single one spreads through everything.

My open-source library genetic_xgb applies that loop to XGBoost, the gradient-boosted trees that suit the tabular data every lock produces, including the physical fingerprint. It keeps a population of candidate models, scores each on data it was not trained on, keeps the best, breeds and mutates them, and stops when the score stops improving. A population of models that rely on different signals is also harder to fool with one forged badge or one cloned voice.

Candidate models, scored on held-out dataBest score so far0.880.920.96Round 1Round 2Round 3Round 4Round 1teal: kept and bred0.810.860.780.900.840.880.800.830.90Round 2teal: kept and bred0.900.880.860.910.870.890.920.850.92Round 3teal: kept and bred0.920.910.900.930.920.940.890.910.94Round 4teal: kept and bred0.940.930.950.930.950.940.960.940.96Illustrative numbers to show the mechanics, not results from a real dataset.
Figure 19. Four rounds of an evolutionary search, one every three seconds. Each bar is a candidate model scored on held-out data. The three best breed the next round, and the best score climbs from 0.90 to 0.96.

Five ways to use it at the physical door

Evolve the fingerprint model
Train on badge, camera, phone, Wi-Fi and door-sensor features, with fitness set to catching borrowed badges at a fixed rate of false holds.
A pickup verification score
Score each pickup or vault release on email domain age, phone history, ID scan, plate and timing, and evolve the model on confirmed fictitious pickups.
A false-alarm filter per site
Each site gets its own population, trained on its own reviewed alarms, because a warehouse at night and a lobby at noon look nothing alike.
Drone versus bird, refreshed
Re-evolve the fusion classifier as new drone types and seasons change the radar and sound signatures.
Shadow mode first
Run every evolved model alongside the current one for a month before it opens or holds a single door.

The signal in the 2026 noise

There were hundreds of security reports in 2026. Six findings held up across the independent sources in this series, and they are the ones I would build next year's plan around.

Exploits are now the main way in
Vulnerability exploitation passed stolen credentials as the top entry point in Verizon's data, at 31% (Verizon), and Mandiant puts exploitation at seven days before the patch on average (Mandiant).
Attackers use AI for scale, not magic
Verizon found threat actors using AI across about 15 attack techniques, with less than 2.5% of them new (SOCRadar on Verizon). JADEPUFFER showed the next step, a whole attack with no person at the keyboard (GCN).
Speed decides
Average breakout fell to 29 minutes (CrowdStrike), while a triage agent handles an alert in about a minute (TechInformed).
Identity is the weak point, online and at the dock
Phishing, social engineering and valid accounts made up 73.3% of initial access attempts (Microsoft), and fictitious pickups rose to about a quarter of cargo incidents (FleetOwner).
AI is now part of the attack surface
AI models or apps were involved in 21% of breaches, and 92% of those had no proper AI access controls (Help Net Security).
Correlation beats any single sensor
Microsoft calls signal correlation the primary strategic variable (Microsoft). Fused sensors at an airport and a physical fingerprint at a vault are the same idea.

Final thoughts

This series followed an attack inward. Apple locks the device in silicon. Exploit prediction tells defenders which bug will be used before it is. Models at the edge, the sign-in page and the help desk decide in seconds, and can themselves be attacked or fail. At the physical door, the same AI forges the identity that opens a loading dock, and the same AI, fused across cameras, radar, Wi-Fi and badge readers, can tell the borrowed badge from the real one.

What the models do

They predict which door will be tried, score every sign-in, request, frame and radar return in seconds, and combine them into one profile that is hard to fake.

What evolution adds

Randomness that keeps attackers from learning the system, a population of models that fail one at a time, staged rollouts and honest labels from reviewed alarms.

The cheapest place to start is the release decision, the moment a load, a box or a bar leaves your control. Log every release with its email, phone, ID and timing signals, label the ones that turn out to be fraud, and evolve a small population of models to score new releases in shadow mode for a month. Then put the best one in front of the person who signs the release, before it holds anything.

Sources

  1. M-Trends 2026: data, insights and strategies from the frontlines, Mandiant, Google Cloud
  2. Cargo theft losses rise 39% as fictitious pickups remain a threat, FleetOwner, November 2026
  3. Cargo crime report, The Loadstar, October 2026
  4. Chaucer backs Lloyd's cargo program targeting strategic theft, Insurance Business, October 2026
  5. Cargo theft report Q1 2026, Insurance Journal, April 2026
  6. The year the rich went wild for gold, Financial Times, August 2026
  7. Demand for private vaults soars in London, Spear's
  8. Louvre thieves escaped with 30 seconds to spare, Senate inquiry, December 2025
  9. One year after Louvre heist, AFP via Kuwait Times, October 2026
  10. IEEE Std 802.11bf-2025, WLAN sensing amendment
  11. IEEE 802.11bf: enabling the widespread adoption of Wi-Fi sensing, NIST
  12. New IEEE Wi-Fi sensing standard and WhoFi research, overview
  13. Motorola Solutions video surveillance strategy, Memoori, 2026
  14. Hikvision 2025 full-year results, SecurityWorldMarket, May 2026
  15. ASSA ABLOY Q4 2025 earnings summary, Quartr
  16. Securitas full year report 2025, Inderes
  17. Securitas Q4 2025 presentation
  18. SECOM company overview FY2026, Envalith
  19. SECOM Q1 FY2026 earnings summary, Quartr
  20. Hanwha Vision 2025 results, Hellot
  21. Hanwha Vision record 2025 results, Boannews, February 2026
  22. dormakaba FY 2025/26 results, wallstreet-online
  23. dormakaba FY 2025/26 press release, finanzen.at
  24. Brink's Q4 2025 earnings summary, Quartr
  25. Loomis full-year report 2025, Inderes
  26. Security industry overview 2025, National Police Agency of Japan
  27. 2025 information security industry survey, Ministry of Science and ICT via Boannews
  28. Physical security market 2026–2034, Fortune Business Insights via GII
  29. Private security services in Switzerland, IBISWorld
  30. Report on private security personnel in Switzerland, Swiss federal government
  31. Microsoft Digital Defense Report 2026, executive summary, October 2026
  32. Microsoft highlights rising cyber risks, Security MEA, October 2026
  33. What the Microsoft Digital Defense Report 2026 means for sign-in, WorkOS
  34. Microsoft report documents first fully autonomous ransomware attack, GCN, October 2026
  35. JADEPUFFER-linked attackers used compromised service principals in Azure, The Hacker News, September 2026
  36. JadePuffer uses AI agents to wipe Azure resources in minutes, Windows Report
  37. JADEPUFFER agentic ransomware, Decryption Digest
  38. The State of Physical Security Operations in 2026, HiveWatch, June 2026
  39. Eyeforce and ABC Supply Co case study, Pelco
  40. Drone disruptions at Copenhagen and Munich airports, Sentrycs
  41. DIDIT response to rising airport drone threats, Telespazio
  42. The drone blockade: airports grapple with a growing threat, Defence Online, November 2025
  43. Drone detection at airports, a&s, 2026
  44. Danish and UK firms rush counter-drone takeover amid airspace breaches, C4ISRNET, December 2025
  45. Germany airport drone disruption management market, MarkNtel Advisors, 2026
  46. Drone threats drive surge in airport defence tech, Reuters via Sowetan, June 2026
  47. Vulnerability exploitation top breach entry point, 2026 DBIR, Verizon, May 19, 2026
  48. 2026 Verizon DBIR: the new era of cyber threats, eSecurity Planet
  49. New Verizon report reveals the security gap attackers are exploiting most, TechRepublic, May 2026
  50. Verizon 2026 DBIR: 10 takeaways, SOCRadar
  51. 62% of breaches involved the human element: key takeaways from Verizon 2026 DBIR, Abnormal
  52. IBM Cost of a Data Breach 2026, Help Net Security, July 30, 2026
  53. IBM 2026 Cost of a Data Breach report: key findings, eSecurity Planet
  54. Shadow AI data breach cost 2026, SureCloud
  55. Shadow AI doubles to 43% of breaches, Forkast
  56. Google's SOC agent cuts alert review from 30 minutes to 60 seconds, TechInformed, June 2026
  57. Microsoft Security Copilot agents, Microsoft
  58. CrowdStrike 2026 Global Threat Report, CrowdStrike
  59. AI now routine in cyber attacks, Google report finds, IT Brief, May 2026
  60. Cloudflare outage on November 18, 2025, Cloudflare
  61. Cloudflare blames Bot Management system for widespread outage, CyberInsider
  62. The file that paralyzed 20% of the Internet, EveryWAN
  63. Google report on adversarial AI use, SiliconANGLE, February 12, 2026
  64. Report: adversarial use of AI is evolving, KnowBe4 on GTIG, May 2026
  65. Google: AI now powers every threat actor's playbook, Cyber Magazine on GTIG's Q3 2026 tracker
  66. CrowdStrike 2026 Global Threat Report findings, CrowdStrike, February 24, 2026
  67. 2026 CrowdStrike Global Threat Report press release, via 01net
  68. How faster cyber-attacks are reshaping cybersecurity strategy, Infosecurity Magazine
  69. CrowdStrike 2023 Global Threat Report, MSSP Alert
  70. Announcing M-Trends 2026, Google Cloud Security Community
  71. FBI report adds AI-enabled scams for the first time, FOX, April 2026
  72. FBI releases annual internet crime complaint report, April 2026
  73. Deepfake statistics 2026, Keepnet, citing Gartner
  74. Anthropic's Project Glasswing finds 10,000+ CVEs in one month, Daily Security Review
  75. AI vulnerability discovery velocity and the disclosure crisis, Cloud Security Alliance, May 2026
  76. Claude Mythos turns years of security research into 20-hour AI exploits, TechRadar
  77. Threat actors are exploiting vulnerabilities faster than ever, TechRepublic on Mandiant's 2023 analysis
  78. Tenable CTO Q&A on AI threats and exposure management, Tenable
  79. The vulnerability curve bent with the AI curve, ProjectDiscovery
  80. Zero-day exploit statistics, DeepStrike, citing Google Threat Intelligence Group
  81. Attackers are handing off access in 22 seconds, Help Net Security
  82. CVSS vs EPSS vs SSVC vs KEV, DefenderMate
  83. Introducing EPSS version 4, Jay Jacobs, Empirical Security, March 17, 2025
  84. Improving vulnerability remediation through better exploit prediction, Jacobs et al., 2019
  85. What is EPSS, SecPod
  86. Google AI Big Sleep stops exploitation of SQLite flaw, The Hacker News, July 2025
  87. Google's new AI doesn't just find vulnerabilities, it rewrites code to patch them, The Hacker News, October 2025
  88. Google DeepMind unveils CodeMender, Ground News coverage summary
  89. Memory Integrity Enforcement, Apple Security Research, September 9, 2025
  90. Operating system integrity, Apple Platform Security
  91. BLASTPASS: NSO Group iPhone zero-click, zero-day exploit captured in the wild, Citizen Lab, September 7, 2023
  92. Apple says it is not aware anyone using Lockdown Mode got hacked, TechCrunch, December 2023
  93. Apple doubles maximum bug bounty to $2M for zero-click RCEs, Security Affairs, October 2025
  94. Security research on Private Cloud Compute, Apple Security Research, October 2024
  95. About Apple threat notifications and protecting against mercenary spyware, Apple Support
  96. Apple warns users in 110 countries of mercenary spyware, TechRepublic, August 2026
  97. Apple notifications alert victims in 99 countries, CyberScoop, October 2026
  98. Apple helps safeguard users from scams and fraud while protecting personal data, Apple Newsroom, October 2025
  99. About Stolen Device Protection for iPhone, Apple Support
  100. iPhone security delay: all you need to know, NordVPN
  101. How Cloudflare auto-mitigated a world record 3.8 Tbps DDoS attack, Cloudflare
  102. Cloudflare thwarts record 7.3 Tbps DDoS attack with automation, SecurityBrief
  103. Cloudflare auto-mitigates 11.5 Tbps attack, SecureBlink
  104. Record-breaking DDoS attack peaks at 22 Tbps and 10 Bpps, SecurityWeek
  105. Cloudflare report shows massive spike in high-volume DDoS attacks, TechRepublic
  106. DDoS attacks over 1 Tbps surged fivefold in the second quarter, BleepingComputer
  107. DDoS attacks hit record scale as 1 Tbps+ campaigns become more common, Help Net Security
  108. Cloudflare reports 18.2 million network-layer DDoS attacks in Q3, The Hacker News
  109. Cloudflare mitigates 23.2 million DDoS attacks, GBHackers
  110. Microsoft Digital Defense Report 2025 summary, Microsoft
  111. Microsoft Digital Defense Report 2025, Libertify summary
  112. Google deploys on-device AI to thwart scams on Chrome and Android, Infosecurity Magazine
  113. How we're using AI to combat the latest scams, Google
  114. Keeping Google Play and Android app ecosystems safe in 2025, Google
  115. Google says its AI systems helped deter Play Store malware in 2025, Reuters via Business Ghana
  116. Train, evaluate, repeat, Michael Manapat, PyData LA 2018
  117. Counterfactual evaluation of machine learning models, Michael Manapat (Stripe), slides
  118. Adversarial machine learning: a taxonomy and terminology of attacks and mitigations, NIST AI 100-2e2025
  119. CrowdStrike shares how a Rapid Response Content update caused global outage, Infosecurity Magazine
  120. CrowdStrike blames outage on content configuration update, Computer Weekly
  121. Disrupting the first reported AI-orchestrated cyber espionage campaign, Anthropic, November 2025
  122. Anthropic disrupts AI-run cyberattack, Decrypt, November 2025
  123. Leigh Van Valen, evolutionary theorist and paleobiology pioneer, University of Chicago News
  124. Self-nonself discrimination in a computer, Forrest, Perelson, Allen and Cherukuri, 1994
  125. genetic_xgb, aleator-labs on GitHub
  126. Hardware wallets, randomness and the vaults behind every device, Zenpanik
  127. Two clocks: quantum breaks the math, AI breaks everything around it, Zenpanik
  128. Nassim Nicholas Taleb, letter to Stewart Brand on fragility and precaution, July 3, 2013
Back to Part 1