Part 3: The AI clock
Part 3 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.
AI attacks the layers around the math
Go back through the incidents in my hardware wallet deep dive: Coldcard's seed bug, eMudhra's key-ceremony script, Bybit's blind approvals, Binance's signing service. None of them broke an elliptic curve. Each one broke something wrapped around it: code, configuration, a ceremony, a person, or a machine allowed to sign. AI makes every one of those layers cheaper and faster to attack, and it needs no new hardware to do it.
- Implementation
- The code that runs the math: TLS libraries, certificate parsers, signing firmware. AI now finds memory-safety and logic bugs in this code faster than human teams can triage them.
- Configuration
- Which algorithms, key sizes and certificates are deployed. Old RSA-1024 keys, forgotten endpoints and weak parameters are what an automated scanner finds first.
- Key lifecycle
- Generation, storage, rotation, backup and ceremonies. These are written procedures run by people and scripts, and each step is a place where a key can leak or weaken.
- People
- The person who can say yes: an approver, a help-desk agent, a KYC reviewer. Deepfakes and AI-written lures go after that person.
- Automation
- Signing services, bots and AI agents that act without a human in the loop. They do what an authenticated caller asks, so whoever reaches the caller controls the key.
Implementation: finding bugs became cheap
On April 7, 2026, Anthropic launched Project Glasswing, giving about 50 partners access to Claude Mythos Preview, a model it has not released to the public. A vulnerability is a flaw in software that an attacker can use to take over a system or read its data. After one month, according to Anthropic, the partners had found more than ten thousand high- or critical-severity vulnerabilities, and Mozilla had fixed 271 Firefox vulnerabilities while testing the model. In October Anthropic reported that partners had found at least 129,000 verified vulnerabilities between April and July, and its own open-source scanning another 5,500; more than 33,000 of the total were rated critical or high.
Cryptographic code is on the list. Anthropic reported that the model found weaknesses in libraries implementing TLS, AES-GCM and SSH that could allow certificate forgery or decryption. TLS is the protocol behind the padlock in a browser, and a certificate is the signed statement that proves a site is who it claims to be. In wolfSSL, a TLS library used in billions of devices, the model built an exploit for CVE-2026-5194 that would let an attacker forge certificates and run a fake website for a bank or an email provider. A signature is computed over a digest, the short fixed-length fingerprint of a document, and the library had skipped checks on the digest's size and on which algorithm it claimed to use. wolfSSL says it fixed eight CVEs from Anthropic's reports in its 5.9.1 release. The RSA and elliptic-curve math inside wolfSSL was intact; the code that checked certificates was not.
Two caveats apply. Small open-weight models found some of the same bugs Anthropic showcased, so attackers do not need Mythos. Few findings get exploited. VulnCheck found that 14 of 1,061 publicly attributed AI-assisted discoveries had been confirmed as used in real attacks, about the same rate as vulnerabilities in general.
Attackers are using the same tools across the whole attack. Anthropic's September 2026 threat report, covering operations it disrupted between December 2025 and August 2026, describes state-linked groups, criminals and lone operators running multi-agent workflows in which the AI did most of the work: scanning exposed services, writing and testing exploits, collecting credentials and sorting stolen data. The report's summary line is that “sophisticated attacks no longer require sophisticated attackers”.
wolfSSL, CVE-2026-5194
An AI model found a signature-verification flaw in a TLS library and built a working certificate-forgery exploit. wolfSSL patched it in release 5.9.1. The cryptographic algorithms were untouched.
RSA-260 and RSA-896
AI coding agents ported a decades-old factoring program to GPUs and pushed the public record from 829 to 896 bits in a month, with no new algorithm. Cognition puts the cost of RSA-1024 at about $30 million. See Part 1.
The speed gap
Attackers also move sooner. The Zero Day Clock, which tracks the gap between public disclosure of a vulnerability and confirmed exploitation, shows the mean falling from 2.3 years in 2018 to under 20 hours in 2026. Critics question how it defines exploitation, and the live figure has moved between hours and days during the year, but every version shows the gap shrinking.
Fixing has not kept pace. Verizon's 2026 Data Breach Investigations Report found that organizations took a median 43 days to remediate known-exploited vulnerabilities, the flaws that attackers are already using, up from 32 days, and fully remediated only 26% of them. As of October 2, Anthropic's disclosure dashboard listed 516 of 6,157 open-source vulnerabilities it had reported to maintainers as patched upstream, to its knowledge.
Anthropic summed up the first month of Glasswing in similar terms. Software security used to be limited by how fast people could find vulnerabilities, and is now limited by how fast they can verify, disclose and patch them. A high- or critical-severity bug found by Mythos Preview took about two weeks to patch on average, and some maintainers asked Anthropic to slow its disclosures because they needed more time to write fixes.
The queue is also filling with noise. On October 1, Google stopped accepting product vulnerability reports to its open-source bug bounty, the program that pays outsiders for reporting flaws, after a rise in automated submissions, most of them invalid. HackerOne's Internet Bug Bounty had paused new submissions in March for a related reason: AI-assisted discovery had outpaced maintainers' ability to ship fixes. Real findings and false ones now compete for the same few reviewers.
Each fix waits on people. Someone triages the finding, assigns an owner, writes and reviews the patch, tests it, schedules a release and deploys it everywhere the code runs. An attacker with the same finding skips every one of those steps.
People: the approver is the new attack surface
Most security processes still end with a person who can say yes, and AI makes that person easier to fool.
US deepfake fraud indictment
On November 5 the Justice Department charged three men with defrauding companies of at least $129 million between July 2024 and February 2026 by using AI to impersonate the voices and images of chief executives and lawyers. Named victims include a UK fintech that lost almost $11.6 million and a Rhode Island bank that wired nearly $2.7 million. The charges are allegations, and one defendant's lawyers say he had no involvement.
Fideuram
In September, Reuters reported that fraudsters took about €95 million from Fideuram, the private banking arm of Intesa Sanpaolo. A spoofed WhatsApp message impersonated the group's CEO, and a phone call in the AI-cloned voice of a senior lawyer served as confirmation. According to Reuters' sources, more than half of the money was later recovered; Intesa Sanpaolo and Fideuram declined to comment.
Deepfake KYC bypass kits
In April a darknet seller reportedly began offering a kit that uses real-time deepfake video and voice to pass identity checks at banks and crypto exchanges. Liveness checks built to catch a photo held up to a camera now face synthetic faces that move in real time.
Zerion
North Korea-linked attackers used AI in a long-running social-engineering campaign against the wallet company and took about $100,000 from its hot wallets, the online wallets used for day-to-day payments. The amount was small, and the same method works against bigger targets.
Arup
A finance worker in Arup's Hong Kong office received a message from the “CFO” about a secret transaction and suspected phishing. A video call with the CFO and colleagues he recognized settled his doubts. Every other participant was a deepfake, and he made 15 transfers totaling about $25.6 million.
In each case the attackers beat a control that assumes people can tell real colleagues from fakes. Any process built on that assumption, from a voice-verified password reset to a video-approved wire transfer, needs a second check that does not depend on recognizing anyone. Defenders can use the same models. At one Glasswing partner bank, Mythos Preview helped stop a fraudulent $1.5 million wire after an attacker took over a customer's email and placed spoofed phone calls.
Automation: when the signer is a machine
Binance's November 3 incident report shows the full chain. In June a wallet DevOps engineer installed a tampered package through Homebrew, a common installer for developer tools on Macs, while looking for an AI coding tool. In September a fake recruiter got a second engineer to run a malicious coding assignment. On October 20 the attackers used that machine and an internal AI tool to take control of two wallet engineers' systems, and found a container deployment token on one of them; the report does not explain how the AI tool was used.
On October 29 they reached the wallet signing service, the internal system that approves withdrawals by signing them with the exchange's keys, and moved about $6.34 billion. The FBI attributed the attack to the North Korean group TraderTraitor, and North Korea denies it. As of November 3, about $2.85 billion had been frozen or intercepted. Every step used access that a compromised person, laptop or tool already had, and one of those tools was internal AI.
Banks have started to restrict outside agents. On October 27, Royal Bank of Canada blocked third-party personal AI agents from its authenticated online banking, citing unacceptable customer risk, and TD and BMO said they restrict them too. Such agents need a customer's login and often store the password to keep working, which turns each one into another holder of credentials with authority to move money.
Agents can also behave in ways nobody asked for. In a July 2026 test run by the UK AI Security Institute, an agent built on Anthropic's Claude Mythos 5 took actions affecting real people and organizations in nine of 43 runs, and OpenAI's GPT-5.6 Sol did so in one run. In the most serious case, the Mythos 5 agent submitted malicious code to an unrelated open-source project and created fake identities to pressure the maintainer, who refused it. AISI found no resulting real-world harm.
The AI clock runs on every ring at once: bugs in code, exposed configurations, faked faces and voices, and signers that act on any authenticated request. Part 4 brings the two clocks together, because the quantum migration is about to add new code and new procedures to every one of these rings.
Continue to Part 4: Where the clocks collide