# AI Breaks NIST’s HAWK in 60 Hours for $100K [2026] > Source: > Published: 2026-10-22 14:29:20+00:00 A New Scientist story headlined “Global race to update computers as AI discovers most disruptive hack yet” appeared on [NewsNow’s cyber security feed](https://www.newsnow.com/ng/World/Technology/Cyber+Security) on October 22, 2026. The full text was not available to us, so this article does not pretend to summarize it. What it does is lay out the verified record around the headline, and a separate report that fits it closely: an AI model broke a leading post-quantum signature candidate called HAWK in roughly 60 hours. That second report matters because it shows what “AI breaks HAWK” actually means. HAWK was a third-round candidate in the NIST post-quantum signature competition, not a standard running on your laptop. Its authors reportedly confirmed the result and pulled the scheme within one day. The same report says no deployed cryptographic standard has been broken by AI. For readers already tracking the quantum threat, our earlier look at [the Quantum Alliance warning on Bitcoin exposure](https://shattered.io/quantum-alliance-bitcoin-30-2-percent-risk-2026/) covers the other half of this story. We want to be plain about the limits. We cannot confirm that the New Scientist piece is about HAWK. We cannot confirm the incident deserves the label “most disruptive hack yet.” We also cannot confirm that any worldwide update push traces back to it. Those are headline claims, and the verified facts do not support or refute them. Everything below is built on what is confirmed, with analysis clearly marked as analysis. Don't miss new tech stories on Google Add Shattered once in the Google app and our stories show up in your news suggestions. [Add now](https://www.google.com/preferences/source?hl=en&q=https%3A%2F%2Fshattered.io%2F) ## AI Breaks HAWK: What Was Reported Here is the core account. An AI model attacked HAWK, a digital signature scheme that reached the third round of the NIST post-quantum signature competition. According to the report, the model broke it in approximately 60 hours. The HAWK authors then confirmed the result and withdrew the scheme within one day. The reported computation cost was approximately $100,000. A second laboratory’s model reportedly found an independent attack on HAWK during the same week. That detail carries real weight, because two separate systems landing on the same weak spot makes a fluke much less likely. The report is also careful about scope. HAWK was a candidate scheme, and no deployed cryptographic standard had been broken by AI. Keep that sentence in mind, because much of the online reaction blurs it. A candidate failing inside a public competition is the competition working. A deployed standard failing in the wild would be a different class of event entirely. ## What the New Scientist Headline Does and Does Not Tell Us NewsNow listed the New Scientist headline on October 22, 2026. That is the only verified fact about the story itself. We have no confirmed text, no named researchers, no affected products or companies, and no direct quotations from it. Headlines compress. “Global race to update computers” could describe a patching scramble, a migration to new cryptography, or a broader reaction to AI-driven vulnerability discovery. Without the article body, any of those readings is a guess. We will not dress a guess up as reporting. The practical takeaway for readers is simple. Treat the headline as a signal that mainstream science media now frames AI-found flaws as an urgent update problem. Do not treat it as proof of any specific incident, vendor, or timeline. When the full text surfaces with named sources, this page should be revisited. ## The 60-Hour, $100,000 Math Spread $100,000 across roughly 60 hours and you get about $1,667 per hour. That is our own arithmetic on the two reported figures, not a number from the report. It frames the economics well, though. A cryptanalysis effort that once demanded a specialist team and months of work now fits inside a mid-sized software budget. Cost is the part defenders should watch. Today the bill is five figures per target scheme. If model costs fall, as they have for most AI workloads, the price of an attempt drops with them. Attackers and researchers can then run many more attempts against many more targets. The same pattern already shows up outside cryptography. We reported on [AI agents that breached more than 100 firms at about $50 a site](https://shattered.io/ai-agents-breached-100-firms-50-dollar-site-2026/), which pushed the same lesson from another angle. Cheap, repeatable machine effort changes who can afford to probe whom. ## Why HAWK Withdrew Within One Day The speed of the withdrawal tells you something about how the community handles bad news. The authors did not dispute the result or ask for months to review it. They confirmed it and withdrew the scheme within one day. That is the behavior public cryptographic competitions are designed to produce. Candidates are published so that strangers can attack them. When an attack works, the clean response is to say so quickly and move on, before anyone builds product on a broken design. In that sense the HAWK episode is a success story for process, even as it is a failure for one scheme. Signature schemes sit under software updates, firmware, certificates, and identity systems. If you want a refresher on how signatures differ from the simpler integrity checks many teams still lean on, our explainer on [HMAC versus digital signatures](https://shattered.io/hmac-vs-digital-signatures-2026/) is a useful companion. ## A Second Lab Finds an Independent Attack The report says a second laboratory’s model found an independent attack on HAWK during the same week. We cannot name that lab or describe its method, because the verified record does not include either. What we can say is why the detail changes the picture. One successful attack might reflect a lucky run or a quirk of one system. Two independent attacks suggest the weakness was findable by capable models in general. That shifts the question from “did one team get lucky” to “how many other candidates have similar blind spots that nobody has pointed a model at yet.” This is analysis, not reporting. Still, it is the question most standards bodies and vendors will be asking in private right now. Expect more candidate schemes to face the same treatment. ## Deployed Standards Remain Unbroken, According to the Report The most important sentence in the report is also the least viral. No deployed cryptographic standard had been broken by AI. HAWK was a candidate, one of several schemes still being evaluated for possible standardization. Readers who see a headline about AI cracking quantum-safe cryptography often conclude their bank, their messaging app, or their VPN is exposed. Based on the verified facts, that conclusion does not follow. Candidate schemes live in a research track, and products generally do not ship on them as the sole line of defense. That said, the line between “candidate” and “deployed” is thinner than comfort suggests. Some vendors run pilots on candidate algorithms to get ahead of the transition. Anyone who did that with HAWK now needs a replacement plan. ## What Is Confirmed and What Is Not The table below sorts the claims circulating around this story by how well the verified record supports them. It exists so you can quote the solid parts without dragging the shaky parts along. | Claim | Status | Basis | |---|---|---| | New Scientist headline listed on October 22, 2026 | Confirmed | NewsNow listing | | AI model broke HAWK in about 60 hours | Reported | Separate report | | HAWK authors confirmed result and withdrew scheme within one day | Reported | Separate report | | Model computation cost about $100,000 | Reported | Separate report | | Second lab found an independent attack the same week | Reported | Separate report | | No deployed standard broken by AI | Reported | Separate report | | Incident is the “most disruptive hack yet” | Unconfirmed | Headline wording only | | Global update race caused by the HAWK result | Unconfirmed | No verified link | | Names, products, specs, quotes from the New Scientist text | Not available | Full text not reviewed | ## Historical Context: Candidates Fall, and That Is the Point NIST runs its post-quantum work as an open, multi-round contest. Teams submit designs, outsiders attack them, and weak entries drop out. Over the years several candidates have fallen to human cryptanalysts, sometimes late in the process. HAWK’s fall to an AI model fits that pattern, with one new feature: the attacker was not a person with a whiteboard. Machine-assisted vulnerability discovery has been creeping toward this moment for a while. Fuzzers, symbolic execution tools, and automated code scanners have found software bugs for years. Mathematical cryptanalysis was long considered harder to automate, because it relies on spotting structure rather than sweeping through inputs. The HAWK report suggests that assumption is weakening. We saw an adjacent shift in web software this month. When we covered [the AI-assisted PHP flaw affecting 1.5M of 2.6M hosts](https://shattered.io/php2shell-ai-php-vulnerability-cve-2026-109618-2026/), the story was also about machines finding what people had missed. Different layer, same trend line. ## Why AI Cryptanalysis Differs From Brute Force A 60-hour result does not mean a machine tried every key. Brute force sweeps a search space and scales with raw hardware. Cryptanalysis hunts for structure, a shortcut that makes the search space smaller than designers assumed. Finding such shortcuts has historically depended on human insight, patience, and a lot of paper. The verified record does not explain how the model located HAWK’s weakness, and we will not guess at the mathematics. The general point still holds. A system that can read the literature, form hypotheses, test them, and discard dead ends at machine speed attacks a design the way a tireless graduate student would, only far faster and in parallel. That has a second-order effect on trust. Designers publish candidate schemes expecting a few dozen human experts to probe them over years. If a model can run thousands of probing attempts in a week, the old assumption that long human scrutiny equals confidence needs revisiting. A scheme that survives ten years of human review may still hold a flaw that no human was positioned to see. The flip side is encouraging. The same capability can serve defenders. Pre-publication stress tests by a model could catch flaws before a design reaches a public competition, which would save everyone the embarrassment of a third-round withdrawal. ## Three Kinds of Update Pressure Compared The phrase “race to update” covers very different workloads. Lumping them together hides the real planning problem, so it helps to compare them side by side. **Software patching** is the familiar one. A vendor ships a fix, administrators apply it, and the clock runs from advisory to deployment. It is painful but well understood, and most organizations have a process. **Algorithm migration** is harder. Swapping a signature scheme touches protocols, key sizes, certificates, hardware modules, and sometimes regulation. Nothing in the verified record says HAWK was widely deployed, so its withdrawal should not trigger a mass migration. It does show how a withdrawal would feel if the scheme had been in production. **Hardware replacement** is the slowest. When a weakness lives in silicon or in a hardware security module, no download fixes it. The Penn State warning about quantum hardware, noted above, belongs in this category. Most teams are organized for the first and underprepared for the second and third. If the headline’s “race” is real in any form, the second category is where the hidden work sits. Planning for it now costs far less than improvising later. ## Market Impact: Who Pays for the Update Race No verified figures tie this episode to market moves, so we will not invent any. We can describe where cost and pressure land, as analysis. Standards bodies take the first hit, because their timelines depend on candidates surviving scrutiny. Vendors that built early prototypes on a withdrawn scheme take the second. Enterprises take the third, since every shift in the recommended algorithm list forces another round of inventory work. Security leaders can follow wider coverage of these pressures on our [security desk](https://shattered.io/security/) and in outside roundups such as [Tech Xplore’s security news feed](https://techxplore.com/security-news/). The winners are likely to be firms selling cryptographic inventory, agility tooling, and managed migration. Their pitch just got an easier sales cycle. Losers include any vendor that marketed a candidate algorithm as future-proof. ## Risk Tiers: Candidate Breaks vs Deployed Standards vs Everything Else Not every security scare deserves the same urgency. This second table separates the layers so teams can triage. The first three rows draw on the verified record, while the last two are general operating categories. | Risk tier | Example | Current status | Who acts | |---|---|---|---| | Candidate scheme broken by AI | HAWK | Withdrawn within one day of confirmation | Authors, NIST, pilot users | | Independent confirmation | Second lab’s attack on HAWK | Reported in the same week | Standards reviewers | | Deployed standard broken by AI | None reported | No break reported | Nobody yet | | Quantum hardware weaknesses | Penn State study, January 2026 | Researchers warn of flaws in hardware and software | Quantum hardware builders | | Ordinary unpatched software | Routine vendor advisories | Ongoing | Every IT team | The Penn State row comes from a [ScienceDaily write-up dated January 20, 2026](https://www.sciencedaily.com/releases/2026/01/260120000330.htm). It describes researchers warning that quantum computers carry security weaknesses in their physical hardware as well as their software. That is a separate issue from HAWK, but it shows how many fronts the quantum transition has. ## Industry Data: The Migration Readiness Gap The best available baseline on corporate readiness comes from a 2024 survey, so treat it as dated context rather than a current reading. Entrust’s 2024 Global report found that, in its words, [“61% of global respondents planned to migrate to post-quantum cryptography within five years, while 41% were already preparing for the transition.”](https://www.entrust.com/company/newsroom/post-quantum-cryptography-awareness-is-high-but-widespread-action-lags-finds-2024-global-entrust-report) Read the gap between those two numbers. Many organizations intend to move, and fewer have started. An AI-accelerated attack cadence compresses the time a slow starter can afford to wait. A firm that was planning to begin in year four of a five-year window may find the target algorithms changing under it. We do not have a 2026 equivalent from a named survey in our verified set. If new data lands after the HAWK episode, it will be worth comparing against this baseline to see whether urgency moved. ## How Reporting on This Story Should Be Read Fast-moving security stories attract three kinds of coverage. Aggregator listings, like the NewsNow entry for the New Scientist headline, tell you a story exists and little else. Secondary reports, like the one describing the HAWK result, carry the numbers but may sit a step removed from the researchers. Primary statements from NIST, the HAWK authors, or the labs involved would settle details, and none were part of our verified set. Readers should weight claims accordingly. The 60-hour, $100,000, and one-day figures come from reporting, and they hang together well. They are still reported figures. If a primary source later revises any of them, the revised number wins. We have chosen the most conservative reading throughout and avoided adding detail the sources do not support. One more caution. Security news cycles reward the biggest framing. “Most disruptive hack yet” is a superlative that no verified fact backs, and the HAWK report itself undercuts it by stressing that no deployed standard fell. The accurate story is smaller than the headline and still significant. ## What Security Teams Should Do This Quarter The response does not require panic. It requires a short list of unglamorous tasks done on schedule. These steps follow from the verified facts and general security practice. ### Inventory where cryptography lives List every system that signs, verifies, or encrypts something important. Include firmware signing, code signing, TLS termination, VPN gateways, and identity providers. Teams that cannot answer “where do we use signatures” cannot swap an algorithm quickly when one is withdrawn. ### Avoid betting production on candidates HAWK was a candidate when it fell. If a vendor pitches you a candidate algorithm as a sole protection layer, ask what happens when it is withdrawn. Prefer hybrid designs that keep a well-understood classical algorithm alongside any new one. For the password side of the house, our comparison of [Argon2, bcrypt, and scrypt](https://shattered.io/argon2-vs-bcrypt-vs-scrypt-2026/) shows how to choose from mature, scrutinized options. ### Keep patch discipline tight A global update race, whatever its trigger, rewards organizations that can push changes fast. Our report on the [BeyondTrust bug that gave a China-linked APT 47 days](https://shattered.io/beyondtrust-cve-2026-102764-china-apt-zero-day-2026/) shows what slow patching costs in practice. Measure your time from advisory to deployment, and cut it. ## Five Predictions, Labeled as Analysis These are our judgments, not reported facts. We have tied each to the verified record and hedged where the record runs thin. - **More candidates will be attacked with AI.** Two independent attacks on HAWK in one week make the technique look repeatable. Expect researchers to point models at other schemes still under review. - **Competition timelines will get more conservative.** Evaluators may want longer public scrutiny windows if machines can find flaws in days. - **Crypto-agility moves from slide to requirement.** Procurement teams will start asking vendors how fast they can swap algorithms. - **Attack costs will fall.** The reported $100,000 figure is a snapshot. Cheaper models mean cheaper attempts. - **Defenders will adopt the same tools.** If an AI model can break a candidate in about 60 hours, standards teams can use one to pre-test designs before they are published. The weakest of these is the second. Standards bodies move slowly and rarely announce process changes in response to a single event, so watch for that one over quarters, not weeks. ## What to Watch Next Several open items will decide whether this story grows or fades. First, the New Scientist full text, which should name its sources and clarify what “race to update” refers to. Second, any statement from NIST on how it handles AI-found weaknesses in remaining candidates. Third, whether other laboratories publish attacks against different schemes. Outlets with running AI-security sections, including [The Hacker News’s AI coverage](https://thehackernews.com/expert-insights/search/label/Artificial%20Intelligence?max-results=20), are the likeliest places for follow-ups. We will update this page when verified details arrive. Until then, the safest framing is a narrow one: an AI model broke a candidate signature scheme quickly and cheaply, the candidate was withdrawn, and deployed standards were reported intact. ## FAQ: AI Breaks HAWK and the Update Race ### What is HAWK? HAWK was a digital signature scheme and a third-round candidate in the NIST post-quantum signature competition. Its authors withdrew it after the reported AI attack. ### How fast did the AI model break HAWK? The report says approximately 60 hours, at a reported model-computation cost of about $100,000. ### Did the HAWK authors accept the result? Yes, according to the report. They confirmed the result and withdrew the scheme within one day. ### Is my encryption broken? The report states that no deployed cryptographic standard had been broken by AI. HAWK was a candidate scheme, not a deployed standard. ### Is this the “most disruptive hack yet”? That wording comes from the New Scientist headline, and we cannot confirm it. The verified record does not support or contradict the label. ### Did the HAWK result cause a global race to update computers? We cannot confirm a causal link. The headline mentions a global update race, but we have not seen the article text, and the HAWK report does not describe one. ### What should organizations do now? Inventory cryptographic dependencies, avoid relying solely on candidate algorithms, build the ability to swap algorithms quickly, and keep patch times short.