Zenpanik
Deep Dive

Two clocks: quantum breaks the math, AI breaks everything around it

Quantum computers will eventually break the cryptography itself. AI is already breaking the processes around it, faster than most organizations can change them.

By Pavel Nikolov, November 11, 2026
Nobody has to pick the lock if the hinges come off.

RSA, the public-key system behind most web certificates and a large part of banking infrastructure, rests on one asymmetry. Multiplying two secret primes p and q into N = p × q takes microseconds, while recovering p and q from N is, as far as anyone knows, extremely slow. Whoever factors N can compute the private key. In 1991 RSA Laboratories published a list of such numbers of growing size as an open test, and the largest one factored so far is the public measure of how far attackers are from the 2,048-bit keys in use today.

In September 2026 that record moved twice in sixteen days. On September 3, Eric Lu of Cognition published a factor of RSA-260, an 862-bit number. On September 19, Steve Weis of Anthropic published both factors of RSA-896. The previous record, the 829-bit RSA-250, took an academic team about 2,700 CPU core-years in 2020, the equivalent of one processor core running for 2,700 years.

Neither team found new mathematics. Both used the general number field sieve, the fastest known factoring method since the 1990s. Its running time grows faster than any polynomial in the number of digits but slower than an exponential, so every digit added to a key multiplies the attacker's work. Lu had Cognition's Devin agent build a GPU version of CADO-NFS, the open-source implementation of the sieve, and run it on about $400,000 of GPU time. Weis had Claude port the same software to GPUs and orchestrate up to 2,048 idle GPUs for ten days, about 30 GPU-years. Weis wrote that there were no new algorithmic factoring improvements and no new threats to deployed keys. Lu estimated that a hyperscaler or frontier AI lab could now factor a 1,024-bit RSA key for about $30 million. Standards retired 1,024-bit keys years ago, but they survive in old devices and certificates, so the estimate puts a price on breaking keys that are still in service.

The two teams left the mathematics where it was and cut the cost of applying it. AI is doing the same to every process wrapped around cryptography. In my last deep dive, on hardware wallets, nearly every large loss came from the process around the key.

Binance lost about $6.34 billion on October 29, 2026. Its November 3 incident report describes a months-long campaign: a tampered software download compromised one wallet engineer in June, a fake job interview compromised another in September, and on October 20 the attackers used an internal AI tool to take over two wallet engineers' machines, one of which held a container deployment token, a credential that lets its holder run software inside the production systems. The report describes no broken cryptography. Every step used access that a compromised person, laptop or tool already had. In this piece I look at two forces now working on cryptography from two directions, at very different speeds.

Quantum attacks the math

Shor's algorithm, published in 1994, factors a number in time that grows only polynomially with its length, but it needs a large quantum computer to run. On such a machine it would recover private keys from public keys for RSA and for elliptic curves. No such machine exists yet. When one does, it breaks the math everywhere at once, including traffic recorded years earlier.

AI attacks everything around the math

AI goes after the code that implements the algorithms, the way systems are configured, the ceremonies in which organizations generate their most important keys, the people who approve transactions, and the machines allowed to sign them. This is happening now, it improves with every model release, and it turns discovery work that took experts months into hours.

I think of them as two clocks. The quantum clock is slow, and nobody knows when it will strike. Each new research paper moves its hands forward, and when it strikes, the damage reaches back in time, because anyone who recorded encrypted traffic years earlier can then decrypt it. The AI clock is fast and already running. It needs no breakthrough. It speeds up with every model release and every tool built on one.

Nassim Taleb's test for fragility separates the two. In his 2013 letter to Stewart Brand, he defines fragility as a nonlinear response: a fragile object suffers more from one large shock than from many small shocks of the same total size, the way a large stone hurts more than the same weight in pebbles. In mathematical terms the harm is a convex function of the shock, so doubling the shock more than doubles the damage. Taleb and Raphaël Douady later defined fragility mathematically as nonlinear sensitivity to volatility. Quantum is the stone, one event that breaks RSA and elliptic curves everywhere at once. AI throws pebbles: thousands of small attacks on code, configuration and people, arriving faster each year. A system can survive the pebbles and still break on the stone, and the reverse holds too.

The two clocks collide during the migration. Defending against quantum means replacing every public-key algorithm in production: new libraries, new certificates, larger keys, hybrid modes that run an old and a new algorithm side by side, new key ceremonies and new hardware firmware. Each of those changes is new code and new process, which is the surface AI attacks. The migration runs on compliance timetables, where validating one cryptographic module, the government certification that banks and agencies require before they can deploy it, often takes 18 months or longer. Exploits now arrive in under a day on average after a vulnerability is disclosed. The gap between those two speeds is several hundred to one.

Processes are never perfect. They need continuous improvement, and the improvements now have to land faster than the attacks. Part 2 explains what quantum breaks and how quickly the estimates are falling. Part 3 maps the layers AI attacks, with recent cases. Part 4 shows where the clocks collide during migration. Part 5 sets out a posture for fintechs, banks and custodians.

Two clocks on one timeline, 2019 to 2026

Every event in this piece, newest first. The tag shows which clock it belongs to: quantum progress, AI-driven attacks, implementation flaws, or the policy deadlines that respond to them. Each name links to its source, and the right-hand link goes to the part that covers it.

Nov 2026AI
US charges $129 million deepfake schemeProsecutors charged three men with using AI to impersonate executives and lawyers to order wire transfers. One defendant denies involvement.
Part 3
Oct 2026Implementation
Compiler leak in ML-KEM and ML-DSALLVM turned constant-time code in a Rust library into a branch on secret data on some targets.
Part 4
Oct 2026Quantum
Alice & Bob: 22,593 qubitsA preprint estimates RSA-2048 could fall in 3.14 days on cat-qubit hardware. The company's largest error-correction chip so far has 12 cat qubits.
Part 2
Oct 2026Policy
NIST IR 8547 finalNIST will deprecate quantum-vulnerable public-key algorithms after 2030 and disallow them after 2035.
Part 2
Oct 2026AI
Binance signing serviceAbout $6.34 billion left hot wallets. A tampered download, a fake job interview and an internal AI tool led attackers to the signing service.
Part 3
Oct 2026AI
Canadian banks block AI agentsRBC, TD and BMO said they bar third-party personal AI agents from logged-in banking.
Part 3
Oct 2026AI
AISI and Anthropic incident analysisIn a July test, an AI agent created fake identities to pressure a real maintainer into accepting malicious code. The maintainer refused.
Part 3
Oct 2026Quantum
Quantinuum: 99.9995% logical fidelityLogical errors fell exponentially as the error-correcting code grew, a step toward fault tolerance.
Part 2
Oct 2026Policy
Europol quantum reportsEuropol named wallet keys with exposed public keys as the main quantum risk and urged early migration.
Part 2
Oct 2026AI
Glasswing resultsPartners found at least 129,000 verified vulnerabilities between April and July 2026 using Claude Mythos models.
Part 3
Sep 2026AI
Fideuram voice-clone fraudReuters reported that fraudsters took about €95 million from Intesa Sanpaolo's private bank with a spoofed message and a cloned voice. More than half was recovered.
Part 3
Sep 2026AI
Anthropic threat reportAttackers ran multi-agent AI workflows across most of the kill chain, from reconnaissance to data theft.
Part 3
Sep 2026AI
RSA-260 and RSA-896 factoredAI coding agents ported a decades-old sieve to GPUs and moved the record 67 bits in sixteen days.
Below
Jul 2026AI
AI finds new math against HAWKAn AI model cut key recovery on a post-quantum candidate's test parameters from 264 to 238 operations; its designers withdrew it.
Part 4
Jun 2026Policy
EO 14412 and OMB M-26-15US federal high-value systems must use post-quantum key establishment by 2030 and post-quantum signatures by 2031.
Part 2
Apr 2026AI
Project Glasswing and CVE-2026-5194An AI model built a certificate-forgery exploit for the wolfSSL library.
Part 3
Apr 2026AI
Deepfake KYC bypass kitA darknet seller reportedly offered real-time face and voice manipulation to get past bank and exchange identity checks.
Part 3
Apr 2026Policy
Cloudflare sets 2029Cloudflare targets full post-quantum security, including authentication, by 2029.
Part 5
Mar 2026Quantum
Google: ECC under 500,000 qubitsBreaking 256-bit elliptic curves needs about 20 times fewer physical qubits than the previous estimate.
Part 2
Mar 2026Policy
Google sets 2029Google set 2029 to finish its own migration and moved authentication up its priority list.
Part 2
Feb 2026Policy
BIP-360 merged as a draftBitcoin's first formal quantum-resistant output type. It does not protect coins already exposed.
Part 2
Jan 2026Policy
G7 financial-sector roadmapThe G7 Cyber Expert Group set out a non-binding post-quantum roadmap for banks, insurers and exchanges, pointing to 2035.
Part 2
Jun 2025Policy
EU coordinated roadmapMember states start by the end of 2026, migrate high-risk systems by 2030 and finish by 2035.
Part 2
May 2025Quantum
Gidney: under a million qubitsThe RSA-2048 estimate fell twentyfold from 2019, with under a week of runtime.
Part 2
Aug 2024Policy
FIPS 203, 204 and 205The first post-quantum standards: ML-KEM for key exchange, ML-DSA and SLH-DSA for signatures.
Part 2
Jun 2024Implementation
KyberSlashTiming leaks in Kyber reference code, the basis of ML-KEM, let attackers recover secret keys.
Part 4
Feb 2024AI
Arup deepfake callA finance worker sent about $25.6 million after a video call in which every other participant was a deepfake.
Part 3
2019Quantum
Gidney and Ekerå: 20 million qubitsThe baseline estimate: RSA-2048 in about eight hours on 20 million noisy qubits.
Part 2
QuantumAIImplementation flawPolicy and standards
Figure 1. Two clocks on one timeline. Twenty-eight events covered in this deep dive, newest first, through November 7, 2026.

AI moved the factoring record without new math

Classical factoring records used to take years of work by small expert teams. RSA-250 fell in February 2020, so six and a half years passed before RSA-260, although CADO-NFS was open source the whole time and GPUs kept getting faster. In September each researcher directed an AI agent to do the engineering: rewrite a mature CPU codebase for GPUs, schedule the work across idle chips, and keep a ten-day run alive. The scarce resource had been the handful of people able to run such a computation, and the agents did most of that work.

512 bits1024 bits1536 bits2048 bitsRSA-250, Feb 2020829 bits, about 2,700 CPU core-yearsRSA-260, Sep 3, 2026862 bits, about $400,000 of GPU time, DevinRSA-896, Sep 19, 2026896 bits, about 30 GPU-years, ClaudeRSA-1024Not factored publicly; about $30 million by Cognition's estimateRSA-2048Quantum's targetBar length = key size. Difficulty grows much faster than length: each extra bit costs more than the last.
Figure 2. The public RSA factoring record by key size. AI agents moved it 67 bits in sixteen days without new mathematics. Sources: Weis, Cognition, Sesame Disk.

Factoring difficulty grows much faster than key length. NIST rates 1,024-bit RSA at about 80 bits of security and 2,048-bit RSA at 112 bits, a gap of 232, or about four billion times more work for the best known attack. So 896 bits is still far from 2048. The cost of applying known methods is what moved. The AI clock works through that same mechanism at every layer: finding implementation bugs, writing exploits, producing convincing fake executives. There is one exception so far, covered in Part 4. In July an AI model found new mathematics against HAWK, a young post-quantum signature candidate.

What each clock threatens

The two threats differ in what they target, when they arrive, how the damage spreads, and who controls the response.

Quantum clockAI clock
What it attacksThe public-key math: RSA, Diffie-Hellman, ECDH, ECDSA, EdDSA, SchnorrEverything around the math: implementations, configurations, key lifecycle, people, automated signers. For young schemes, the math as well (Part 4)
WhenNo capable machine yet. Expert views run from “cannot rule out 2029” to “unlikely before 2035” (Part 2)Now, and faster with each model release
How damage spreadsAll at once and backwards in time: recorded traffic and exposed public keys fall togetherContinuously, one target at a time: a bug, an approval, a signer
DefenseA migration: new algorithms, new keys, new certificatesA loop: find, fix, verify and repeat, faster than the attacker
Who sets the paceStandards bodies and regulatorsYour own engineering and operations teams

Quantum readiness fits a project plan with a budget and an end date. The AI clock has no end date, and attackers exploit the habits that big migration projects create: long change freezes, batched releases, and approvals that take longer than the exploit.

Continue to Part 2: The quantum clock