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

Part 4: Where the clocks collide

Part 4 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.

Migration is the attack surface

Moving to post-quantum cryptography touches every TLS endpoint, certificate authority, hardware security module (HSM, the tamper-resistant box that holds an organization's keys), payment terminal, mobile app, API gateway, firmware signing pipeline and wallet. For each one, a team has to choose a library, rebuild, reissue keys and certificates, replace hardware that cannot be patched, and test that nothing breaks.

Every one of those steps lands in the outer rings from Part 3. New libraries are new implementations, new parameters and hybrid modes are new configuration, new keys need new ceremonies, and new approval flows bring new people and new automation. The quantum clock forces the change, and AI is now the cheapest way to find mistakes in it.

The wolfSSL bug from Part 3 shows the collision inside one CVE. According to NVD, missing digest-size and algorithm-identifier checks weakened ECDSA certificate verification in builds that also enabled EdDSA or the post-quantum ML-DSA. A flaw in classical signature checking sat in exactly the configurations that had started to add new algorithms.

The new algorithms are also much larger, and size affects packets, certificate chains, smart-card memory, blockchain fees and the buffers that parsers have to handle. Code that meets new sizes for the first time is where bugs appear.

Signature size, bytesEd2551964ECDSA P-25664ML-DSA-442,420ML-DSA-653,309SLH-DSA-SHA2-128s7,856Key exchange, bytes sent (client, then server)X2551932 + 32ML-KEM-7681,184 + 1,088
Figure 9. An ML-DSA-65 signature is about fifty times the size of an Ed25519 signature. Sizes from FIPS 204, FIPS 205 and FIPS 203.

Dan Boneh of Stanford, a co-author of Google's March paper, said in an interview with Isabel Foxen Duke that a hasty move to post-quantum cryptography is “more likely to cause a catastrophic bug” than a quantum attack. He argued that Bitcoin should consider hybrid signatures that combine elliptic curves with post-quantum schemes instead of a single jump.

Taleb gives the same warning a statistical form. In his letter to Stewart Brand he writes that time has a lot of value in detecting what is breakable and what is not, and ties this to the Lindy effect: the longer something has survived, the more evidence we have that it will keep surviving. RSA has survived nearly fifty years of public attack. The post-quantum algorithms have had about a decade of open review, and that review has already removed candidates. SIKE reached the fourth round of NIST's process in 2022 and was then broken by Wouter Castryck and Thomas Decru on a classical computer.

2023 to 2024

KyberSlash

A timing side channel leaks a secret through how long a computation takes, which is why cryptographic code is written to run in the same time whatever the secret is. Kyber's reference code, the basis of ML-KEM, used division instructions whose timing depended on secret data. In demonstrations, attackers recovered secret keys within minutes for one variant and hours for the other. A team writing a formally verified Kyber in Rust noticed the flaw first, and libraries patched it quickly. The algorithm was sound; the code leaked.

Follow-up

Power analysis on patched ML-KEM

A later study showed the fixed implementations still leaked through power consumption and recovered the secret key in about 30 seconds on a Cortex-M4 microcontroller, the class of chip in many payment and IoT devices. Closing one side channel moved the leak to another.

2026

Compiler and fault attacks on ML-KEM and ML-DSA

On October 30, the Rust library orion patched a leak that its own source code did not contain: on targets without conditional-move instructions, the LLVM compiler turned a helper used throughout its ML-KEM and ML-DSA arithmetic into a jump on secret data. In March, CVE-2026-3503 described how an attacker with physical access could glitch an ARM Cortex-M chip so that it miscalculates, and use the induced faults to compromise key material in wolfSSL's ML-KEM and ML-DSA. The algorithms held in both cases; the build toolchain and the hardware did not.

2026

eMudhra key ceremony

From my last deep dive: a faulty script in a certificate authority's key ceremony produced intermediate CA keys that could be recovered from the public certificates in under a second. Every post-quantum certificate hierarchy will need new ceremonies, run with new scripts.

Stateful signatures: a process risk by design

CNSA 2.0 requires software and firmware signing to use CNSA 2.0 algorithms exclusively by 2030, and for that use case it specifies LMS and XMSS, the stateful hash-based signatures in SP 800-208. Their security rests only on hash functions, the part of cryptography quantum computers barely touch.

Their weakness is operational. Each signing key is a large set of one-time keys, and each may be used only once. If the signer's state is restored from a backup, copied to a second HSM for redundancy, or rolled back after a failed release, the same one-time key can sign twice. Two signatures from one one-time key reveal enough of it for an attacker to forge others. The mathematics is quantum-safe. Whether it stays safe depends on an operational process, the layer AI-equipped attackers probe.

Assumptions under parallel search

Every cryptographic standard rests on assumptions. Some are mathematical: no efficient algorithm exists for factoring, for the lattice problems behind ML-KEM, or for telling AES apart from a random permutation. Others are economic and social: an attack that needs 264 operations, about 1.8 × 1019, and a rare specialist will not happen, and two years of expert review without a break counts as evidence of strength. A model that writes code, runs experiments and tests many hypotheses in parallel puts pressure on all of these assumptions, though not equally, and the evidence so far cuts both ways.

In July 2026 the pressure reached the mathematics. Anthropic reported that Claude Mythos Preview, working in a semi-autonomous multi-agent setup with occasional human project management, found a new key-recovery attack on HAWK in about 60 hours and roughly $100,000 of API cost. HAWK, a lattice-based signature scheme, had survived two years of review in NIST's additional-signature process. The attack uses a previously unused symmetry in HAWK's structure and cuts the work to recover a key on the HAWK-256 challenge parameters from about 264 to 238; a key was recovered end to end in hours on one server. That is a reduction by a factor of 226, about 67 million. The challenge parameters are a deliberately small version published so that researchers can test attacks; the sizes proposed for real use are larger. The HAWK team confirmed the result and withdrew the scheme from NIST's process the next day.

The same results show the limits. The HAWK attack is still exponential: HAWK-512 still needs about 2108 operations, and Anthropic says the attack does not extend to Falcon, ML-DSA or other lattice schemes. The second result applies to a seven-round version of AES. The real cipher scrambles data in ten rounds, and cryptanalysts attack versions with fewer rounds to measure how much margin the full cipher has. PostQuantum.com's analysis puts the gain at a constant factor over a 2013 attack, and the attack needs about 2105 chosen plaintexts, messages the attacker picks and has encrypted under the secret key, which Anthropic itself calls completely impractical. Checking the AES result took several hundred researcher-hours.

On CryptanalysisBench, a 191-task benchmark from ETH Zurich, Anthropic, Tel Aviv University, the University of Haifa and TU Berlin, models broke 6 to 12 of 142 schemes that have no known practical break at full strength. Two of those breaks were new: a full key recovery on the SpoC authenticated cipher, found independently by two models, and an error in the published security proof of KINDI. Earlier machine-learning attacks on lattices, such as Meta's SALSA, recovered secrets only for small dimensions with sparse binary secrets, far from the parameters in ML-KEM.

These cases show where parallel hypothesis testing works. A model can try thousands of ideas when each one is cheap to check. A fuzzer, which feeds a program millions of random inputs to see which ones crash it, answers in milliseconds. A candidate factor either multiplies back to the target or it does not. A key-recovery idea can be tried on a toy-size version of a cipher. For AES, the team proved the key step, formalized part of the proof in Lean, a proof assistant that checks mathematics by machine, and recovered keys on smaller AES-like ciphers, but never ran the full attack. AI erodes an assumption fastest when the assumption is cheap to test.

Serial reviewone line of inquiry at a timeHAWK: two years of expert reviewwithout finding the symmetryParallel model workflowmany hypotheses, each with a cheap checkProof and human verificationHAWK: about 60 hours to find. AES: several hundred researcher-hours to check.Fast where each idea is cheap to test: fuzz inputs, candidate factors, toy-size ciphers. Full-size hardness cannot be tested this way.
Figure 10. Schematic, animated. Parallel search finds candidates quickly; checking them stays slow. Amber dots are hypotheses failing their checks, the dark dot a candidate that survives. Sources: Anthropic, ForkLog.
AssumptionCheap to test in parallel?Evidence so farStatus
The code does what the specification says (constant time, correct checks)Yes: run it, time it, fuzz itwolfSSL CVE-2026-5194, KyberSlash, the LLVM leak in orionFailing now
Attacks cost more than attackers will payYes, on scaled-down instancesRSA-896 in ten days; RSA-1024 at about $30 million (Part 1)Failing for older parameters
Years of expert review mean no shortcut existsPartly: toy parameters first, then proofsHAWK withdrawn after about 60 hours of model workWeakened for young schemes
The underlying problem is hard at standard sizesNo: full-size tests are infeasibleNo standardized primitive affected; AES gain limited to 7 roundsHolding
People can recognize colleagues; signers check intentYes: attackers iterate on targetsArup, Fideuram, Binance (Part 3)Failing now

Taleb warns in his letter against treating absence of evidence as evidence of absence, and puts the burden of evidence on whoever introduces something new. Applied here, the absence of a known attack is weaker evidence than it looks. The results also follow his Lindy logic. HAWK, two years into review, lost 26 of its 64 bits of security on its challenge parameters. AES, standardized in 2001, gave up a constant factor on a reduced version. Old, heavily attacked primitives have earned more trust, and the youngest post-quantum schemes deserve hybrid protection while models probe them. On NIST's forum, Markku-Juhani Saarinen argued that AI-assisted cryptanalysis should ship machine-checkable proofs, which moves the bottleneck from Part 3, verification, into mathematics.

Two speeds that do not match

Validating a cryptographic module under FIPS 140-3, the US and Canadian government standard that regulated buyers require, often takes 18 months or longer, and FIPS 140-2 certificates moved to NIST's historical list on September 21, 2026, so many vendors are in that queue now. Organizations took a median 43 days to remediate known-exploited vulnerabilities in Verizon's 2026 data. A high- or critical-severity bug found by Mythos Preview took about two weeks to patch. Attackers exploit disclosed bugs in under a day on average.

1 hour1 day1 week1 month1 yearFIPS 140-3 module validation18 months or moreKnown-exploited bug, median fix43 daysAI-found bug, average patchabout two weeksDisclosure to exploitation, 2026under a dayLog scale. Each bar is from a different source and population; read them as orders of magnitude.
Figure 11. The cadence ladder. Certification and typical remediation run in weeks to years (ink); dedicated patching of AI-found bugs takes about two weeks (teal); exploitation takes hours (amber). Sources: Encryption Consulting, Verizon DBIR 2026 via watchTowr, Anthropic, Zero Day Clock via Splunk.

Taleb's letter explains why the slow lane loses. He argues that a decentralized system beats a command-and-control, bureaucratic one because its “errors are decentralized and do not spread”. In his picture, nature produces many small, independent variations, while human systems produce fewer and larger ones, and a single large deviation can come to dominate the total. A migration run as one centrally scheduled program concentrates change into large releases, each a large deviation. Small, frequent, independently deployable changes keep each error local.

A compliance-led program, with an inventory spreadsheet, a steering committee, a multi-year roadmap and an annual audit, fits the quantum clock, which has a rough date and a known fix. It fails against the AI clock, because each step of the migration releases new code and new procedures that attackers can probe within days. OMB's own memo tells federal agencies that manual processes are often inadequate for a migration of this scope. Certification shows that a module was built correctly at one point in time. It cannot show whether the deployment around it can be fixed this week.

A Taleb lens, revisited

In the previous deep dive I used Taleb's coffee cup: the atoms move at random, the cup stays put, and the risk is someone knocking it off the desk. The two clocks extend the picture. Quantum is a large stone with an unknown arrival date. For harvested data it is also irreversible, because once a recording is decrypted nothing makes it secret again. Taleb warns about this kind of irreversible risk, which is noticed only after the damage is done. AI is a stream of pebbles with the occasional black swan, such as Binance.

The fragile response to quantum is a big-bang migration close to the deadline. The fragile response to AI is a slow change process that batches fixes. The same design works against both clocks: small, frequent, reversible and verified changes, so each shock stays local and the process learns from it.

What breaks: math at top, process below

Slow threat to the math

Quantum computers and harvest now, decrypt later. A known fix on a long, uncertain timetable.

Fast threat to the math

AI-assisted classical cryptanalysis: known attacks get cheaper. RSA-896 in ten days; RSA-1024 estimated at about $30 million.

Slow defense of the process

Compliance-led migration: multi-year roadmaps, 18-month certifications, 43-day median fixes. The clocks collide here.

Fast threat to the process

AI-found bugs, deepfaked approvers, scanned configurations, signers that act on any authenticated request. Exploits within a day of disclosure.

Speed: slow on the left, fast on the right
Figure 12. Four quadrants of the problem. Organizations control the outlined quadrant, and the fast quadrant beside it attacks it.

No fintech, bank or custodian controls the pace of quantum hardware or of AI. Each one controls how quickly its own cryptography, configuration and approval rules can change. Part 5 turns that into a posture.

Continue to Part 5: Posture at machine speed