Zenpanik
Two clocks: quantum breaks the math, AI breaks everything around itBy Pavel Nikolov, November 11, 2026

Part 5: Posture at machine speed

Part 5 of 5 in a deep dive on how quantum computing and AI are breaking cryptography from two directions. Start with Part 1 if you are new to the series.

Two lanes

Some cryptographic decisions should stay slow and deliberate: which algorithms and parameters to adopt, how to run a root key ceremony, when to change HSM firmware. Boneh's warning about rushed migrations applies here, because a mistake in this lane is expensive and hard to undo. Everything else, from patching a library to tightening an approval rule, should move continuously.

I would split the work into two lanes with different rules. Governance decides what belongs in each lane and who may approve changes in it, and leaves the speed of the fast lane to the engineering teams.

Deliberate lane

Rare, high-impact, hard to reverse
Choose algorithms and parametersyearly review
Root and CA key ceremoniesper event
HSM and secure-element firmwarequarters
Stateful signing key managementper event
On-chain protocol changesyears
Formal review, independent sign-off, rehearsed rollback. The goal is to get each change right once.

Continuous lane

Frequent, small, reversible
Patch cryptographic librarieshours to days
Rotate leaf certificates and API keysautomatic
Find and remove weak or legacy keyscontinuous
Update approval and fraud rulesdays
Retire exposed addresses and endpointsdays
Pre-approved change paths, automated tests, staged rollout, instant rollback. In this lane speed is the control.
Figure 13. A two-speed operating model for cryptography, as I propose it. The deliberate lane guards decisions that are hard to undo; the continuous lane absorbs what the AI clock finds. The moving dots show relative pace. The cadences are targets, not measurements.

Seven practices for both clocks

1. Inventory continuously and remove first

Every migration plan starts with knowing where cryptography lives. OMB's M-26-15 calls the cryptographic inventory the foundation of an agency's plan and expects it to be continuously updated, with automated discovery. Rebuild yours from code, certificates, cloud configuration and HSM logs, and use AI-assisted scanning to find every call to RSA or elliptic curves, every hard-coded key and every RSA-1024 certificate still in service.

Then apply Taleb's via negativa. In his letter he argues that the effects of adding something new are hard to predict, while the effect of removing something that was not historically part of a system can be predicted. Removing RSA-1024 keys, unused certificates, reused addresses and standing signing permissions shrinks the attack surface with no new code. Do that before adding anything.

2. Make algorithms configuration, not code

Crypto-agility means that swapping an algorithm, a parameter set or a broken library is a configuration change and a redeploy. Route cryptographic operations through one internal interface, version your key and certificate formats, and test the swap path before you need it. The first post-quantum implementations will have bugs, as KyberSlash showed, and agility is how you replace them in days.

3. Hybrid key exchange now, signatures next

Harvest now, decrypt later threatens confidentiality today, so key exchange moves first, as a hybrid of X25519, today's standard elliptic-curve key exchange, and ML-KEM. A hybrid stays secure if either half holds. NIST's IR 8547 supports hybrid schemes during the transition, and Cloudflare reports that over 65% of human traffic to its network is already post-quantum encrypted.

Signatures come next. Google and Cloudflare both target 2029 and have moved authentication up their priority lists, because forged signatures would let an attacker impersonate trusted services. Start with the longest-lived keys: roots, firmware signing, and anything whose public key is published permanently. Keep the classical half of the hybrid for as long as the standards allow. In July a young post-quantum candidate, HAWK, lost 26 of its 64 bits of security on its challenge parameters to about 60 hours of model work.

4. Shrink quantum exposure in custody

For anyone holding digital assets, the cheapest quantum defense is hygiene. Google's paper recommends that holders stop exposing and reusing vulnerable addresses. Move funds out of outputs whose public keys are on-chain, know which institutional wallets have exposed keys, and plan their move to quantum-resistant outputs once the protocols allow it. Europol reached the same conclusion in October. For keys already exposed, it found pre-emptive migration to be the only remedy. A hardware wallet protects the private key; only the way you use addresses protects the public one.

5. Verify intent, not identity

Deepfakes break any control built on recognizing a face or a voice. The Arup worker recognized his colleagues on the call, and he moved $25.6 million because of it. The Fideuram fraud and the $129 million US indictment used the same pattern: a written request in a senior person's name, then a call in an AI-cloned voice to confirm it. Replace recognition with checks an impersonator cannot pass: call-backs to numbers held on file, quorum approval for large transfers, limits that one person cannot raise, and clear signing that shows the approver what will happen. This is the intent-safety layer from the hardware wallet deep dive, extended to every human approval.

6. Point AI at your own code first

The same class of models attackers use can scan your stack before they do. Anthropic reports that enterprises using its Claude Security tool, built on the generally available Claude Opus 4.7, patched over 2,100 vulnerabilities in three weeks. Fund the fix pipeline along with the finding pipeline: an owner per finding, an agreed fix time by severity, and a fast lane to production.

The same models can also prove properties of code. In October, Anthropic and the verification company Theorem reported that Claude produced machine-checked proofs that 239 fuzz-test entry points in 123 open-source libraries cannot crash on inputs under 64 MB. A fuzz-test entry point is a function that automated testing feeds with random inputs. The proofs hold at the level of compiled RISC-V code, and the work found hundreds of bugs on the way. KyberSlash was first noticed by a team writing a formally verified Kyber. Anthropic expects better models and reusable proofs to keep lowering the cost of this kind of verification.

7. Split authority and keep humans on irreversible steps

Taleb's first precautionary rule in the letter is to split sources of harm, because harm grows faster than size. At Binance, attackers who reached one internal signing service moved $6.34 billion. Let automation handle detection, low-risk changes and routine rotation. Keep a human quorum on the actions that cannot be undone: moving funds above a threshold, operating root keys, changing approval policy, and granting signing authority to an AI agent. Anthropic's threat report recommends treating AI keys and agent integrations as production credentials. At Binance an internal AI tool was one step on the path to the signing service. RBC, TD and BMO now bar outside agents from logged-in banking, and CIBC tells clients not to share banking credentials with them.

LayerPost-quantum migrationContinuous AI scanningCrypto-agilityIntent verificationHuman quorum
Math
Implementation
Configuration
Key lifecycle
People
Automation
StrongPartialNot covered
Figure 14. Coverage of the six layers by five controls, as I assess it in this piece. Post-quantum migration covers the core; the other four controls cover the rings the AI clock attacks.

A post-quantum migration on its own protects the math and does nothing for the five layers around it.

Where to start

Math safety

Build the inventory and remove what you no longer need. Turn on hybrid key exchange wherever your stack supports it. Plan signature migration for roots and firmware first. Move custody funds out of outputs with exposed public keys.

Process speed

Measure the time from disclosure to a deployed patch for your cryptographic libraries, set a target in days, and give that lane pre-approved changes. Put a human quorum on every irreversible action, including any signing authority granted to an AI agent.

Sources

  1. Factoring RSA-260, Cognition
  2. RSA-896, Stephen A. Weis
  3. AI and classical factoring security, Sesame Disk
  4. SP 800-57 Part 1 Rev. 5, Recommendation for Key Management, NIST
  5. Steve Weis on RSA-896, X, September 20, 2026
  6. Hardware wallets, randomness, and the vaults behind every device, Zenpanik
  7. Rigged coding assignment, abused AI tool opened path to $6 billion Binance heist, SecurityWeek
  8. Nassim Nicholas Taleb, letter to Stewart Brand, Artangel Longplayer Letters, July 3, 2013
  9. Nassim Nicholas Taleb, On the difference between a large stone and a thousand pebbles, Antifragile draft chapter
  10. Taleb and Douady, Mathematical Definition, Mapping, and Detection of (Anti)fragility, Quantitative Finance, 2013
  11. PQC migration: the hard realities, Encryption Consulting
  12. The agentic SOC will help you win the AI era, Splunk
  13. Authorities accuse trio of using AI deepfakes to defraud $129M from businesses, CyberScoop
  14. RUSTSEC-2026-0426: LLVM timing leak in orion ML-KEM and ML-DSA, Wiz vulnerability database
  15. Pappalardo et al., Shor's algorithm compiled on cat qubits: 22 593 qubits for RSA-2048 in 3.14 days, IACR ePrint 2026/2665
  16. NIST starts the countdown to retire quantum-vulnerable encryption, Help Net Security
  17. Binance says fake job interview led to $6.3 billion crypto heist, BleepingComputer
  18. Canadian banks restrict third-party AI agents, Insurance Journal (Bloomberg)
  19. Claude Mythos 5 used fake accounts to pressure a real GitHub maintainer, The Next Web
  20. Quantinuum breaks error-correction world record by 10x, Quantinuum
  21. Europol warns crypto wallets are the primary risk for quantum attacks, Decrypt
  22. Expanding the Cyber Verification Program, Anthropic
  23. AI messaging scam costs Italy's top bank Intesa millions, sources say, Reuters
  24. Detecting and countering misuse of AI: September 2026, Anthropic
  25. Discovering cryptographic weaknesses with Claude, Anthropic
  26. OMB M-26-15, Execution of the Migration to Post-Quantum Cryptography, The White House
  27. Project Glasswing: an initial update, Anthropic
  28. New AI cybercrime tool targets crypto and bank KYC systems via deepfakes, Cointelegraph
  29. Cloudflare targets 2029 for full post-quantum security, Cloudflare
  30. Safeguarding cryptocurrency by disclosing quantum vulnerabilities responsibly, Google Research
  31. Quantum frontiers may be closer than they appear, Google
  32. BIP-360: Pay-to-Merkle-Root (P2MR), Bitcoin BIPs repository
  33. G7 Cyber Expert Group roadmap for the transition to post-quantum cryptography in the financial sector, US Treasury
  34. A Coordinated Implementation Roadmap for the Transition to Post-Quantum Cryptography, European Commission
  35. Craig Gidney, How to factor 2048 bit RSA integers with less than a million noisy qubits, arXiv:2505.15917
  36. FIPS 203, Module-Lattice-Based Key-Encapsulation Mechanism Standard, NIST
  37. Bernstein et al., KyberSlash, IACR ePrint 2024/1049
  38. British engineering giant Arup revealed as $25 million deepfake scam victim, CNN
  39. FIPS 204, Module-Lattice-Based Digital Signature Standard, NIST
  40. FIPS 205, Stateless Hash-Based Digital Signature Standard, NIST
  41. SP 800-208, Recommendation for Stateful Hash-Based Signature Schemes, NIST
  42. NIST announces HQC as fifth standardized post-quantum algorithm, SecurityWeek
  43. The Black Swan: The Impact of the Highly Improbable, Wikipedia
  44. Alice & Bob estimates 22,593 qubits could break RSA-2048, PostQuantum.com
  45. How many qubits does it take to break RSA and ECC?, PostQuantumSecurity
  46. Sandia, Quantinuum and NVIDIA introduce QUOPS, Quantum Computing Report
  47. Scientists unveil Helios, Live Science
  48. When will quantum computers break RSA-2048?, Q-Day Is Coming
  49. Bitcoin bulls scramble for post-quantum protection, CoinDesk
  50. Stanford cryptographer warns against rushed transition, crypto.news
  51. Michele Mosca, Cybersecurity in an era with quantum computers: will we be ready?, IACR ePrint 2015/1075
  52. Google shortens timeline for quantum-safe encryption transition, The Quantum Insider
  53. Quantum Alliance counts 6.04 million bitcoin with exposed public keys, CryptoTicker
  54. BIP-361: Post Quantum Migration and Legacy Signature Sunset
  55. Europol quantum reports on cryptocurrency and harvest-now-decrypt-later, PostQuantum.com
  56. NSA CNSA 2.0 FAQ update, PostQuantum.com
  57. The US federal PQC mandate, 2026, PostQuantum.com
  58. G7 Cyber Expert Group releases roadmap, Orrick InfoBytes
  59. Project Glasswing, Anthropic
  60. Anthropic's new AI model finds and exploits zero-days across every major OS and browser, Help Net Security
  61. CVE-2026-5194, National Vulnerability Database
  62. How Claude Mythos Preview helped harden wolfSSL, wolfSSL
  63. Small open models hunt the same bugs Anthropic showcased, The Decoder
  64. Just 1% of AI-discovered vulnerabilities exploited in the wild, Infosecurity Magazine
  65. AI lowers barriers for sophisticated cyberattacks, The Tribune
  66. Zero Day Clock
  67. Zero Day Clock: all the pieces matter, Jericho
  68. Data Breach Investigations Report 2026, Verizon
  69. Verizon DBIR 2026: vulnerability exploitation is now the top way breaches start, watchTowr
  70. Anthropic's Cyber Mission starts with 6,157 findings reported to maintainers and 516 patched, VentureBeat
  71. Zero Day Clock annual figures, AI-Era
  72. Google narrows open-source bug bounty amid wave of invalid automated reports, SecurityWeek
  73. Italian bank loses $108M to AI messaging fraud, Biometric Update
  74. Phishing, deepfakes to fuel 2026's biggest crypto hacks, Cointelegraph
  75. North Korea denies role in $6 billion Binance hack after FBI accusation, CryptoSlate
  76. Investigating alignment failures in AISI's incident, UK AI Security Institute
  77. NIST Post-Quantum Cryptography Standardization, Wikipedia
  78. KyberSlash timeline and papers
  79. Power analysis attack on patched ML-KEM implementations, IACR
  80. CVE-2026-3503: fault injection attack with ML-DSA and ML-KEM on ARM, CVEFeed
  81. eMudhra certificates private keys compromised, Help Net Security
  82. Trimming wolfCrypt to PQC for DO-178, wolfSSL
  83. AI broke a NIST candidate. Not your encryption, PostQuantum.com
  84. Claude AI cracked a post-quantum test scheme and found a faster 7-round AES attack, The Hacker News
  85. Claude Mythos Preview helps Anthropic find weaknesses in HAWK and 7-round AES-128, ForkLog
  86. CryptanalysisBench: Can LLMs do Cryptanalysis?, arXiv:2607.18538
  87. Researchers publish CryptanalysisBench to verify AI-generated cryptographic attacks, RuntimeWire
  88. Wenger et al., SALSA: Attacking Lattice Cryptography with Transformers, IACR ePrint 2022/935
  89. Anthropic posts AI-found HAWK attack to NIST forum, Implicator
  90. PQC for government and defense, Encryption Consulting
  91. Hardware wallets deep dive, Part 2: A Taleb lens, Zenpanik
  92. NIST IR 8547 explained, Encryption Consulting
  93. Hardware wallets deep dive, Part 5, Zenpanik
  94. LLMs are making formal software verification feasible at a much larger scale, Anthropic
  95. Some Canadian banks block third-party AI agents, Global News
Back to Part 1