Claude, you're starting on the cryptographic algorithms.
Note: This is not about finding bugs in the code, but about identifying flaws in the mathematics itself.
Today, Anthropic released a groundbreaking study: Claude Mythos Preview discovered entirely new attack methods on two cryptographic algorithms.
The first target was HAWK—a candidate in the U.S. NIST post-quantum signature competition that has already withstood two years and two rounds of review by the world’s top human experts. Mythos spent 60 hours and directly halved its effective key strength.
The second goal is even tougher: AES, the most widely used symmetric encryption algorithm in the world—a "hard nut" that humanity has studied for 25 years.
Mythos developed a novel attack method on the simplified AES-128, naming it "Möbius Bridge," which accelerates the previous strongest attack by a factor of 200 to 800.
Two achievements, each burning approximately $100,000 in API fees.
A year ago, language models couldn't even perform basic cryptanalysis. A year later, they found flaws in a scheme that had been reviewed by humans for two years—and even named their own discoveries.
Even on Twitter, some users are worried about their Bitcoin.
60 hours to unlock a lock that had been under review for two years
When Claude Mythos Preview was first released, it was already able to find vulnerabilities in nearly all software, including several major cryptographic libraries.
But those vulnerabilities, in plain terms, belong to the "implementation" level—programmers misapplied the algorithm, wrote flawed logic, or failed to consider edge cases. Such bugs can be found by even an experienced security engineer.
But this time is different. It found a mathematical flaw in the algorithm itself.
For example: Previously, we were trying to find out who installed the lock crookedly; now, we need to prove that the lock’s design principle itself is flawed.
What is HAWK? It is a leading contender in NIST’s Post-Quantum Cryptography Standardization Competition. In 2022, the U.S. National Institute of Standards and Technology issued a global call: Quantum computers are inevitable, and existing classical signature algorithms like RSA and ECDSA will be shattered by quantum computing power—who can create a new lock that even a quantum computer cannot break?
HAWK was one of the applicants and advanced all the way to the third round, surviving two years of intensive scrutiny by global cryptography experts.
Its security is based on the mathematical hard problem known as the Lattice Isomorphism Problem.
In simple terms: HAWK’s key is hidden inside a massive high-dimensional lattice, and to crack it, you’d need to find a hidden path within that lattice. Mathematicians think finding that path is so difficult that they’ve confidently built their security on it.
What did Mythos do? It discovered a previously unused symmetry in the grid used by HAWK—known in academic terms as a "nontrivial automorphism."
Previous papers have proven that if such symmetry can be efficiently found, it can be exploited for an attack. However, that was merely a theorem, and no one has ever been able to determine whether HAWK's lattice actually possesses this symmetry.
Mythos answered: Yes.
The key strength is directly halved. The smallest HAWK-256 parameter set, which was originally expected to require a cost of 2^64—approximately 180 quintillion operations—for full key recovery.
Mythos proof requires only 2^38, approximately 270 billion operations—a reduction of about 4.4 trillion times.
If HAWK wants to maintain its original security level, the key size must double. But once the key size doubles, it loses its core advantage as a "lightweight" post-quantum candidate—small signature size and fast speed.
The most intriguing aspect of this process is that the human operators behind the project are not lattice cryptography experts.
He has a background in theoretical computer science but is unfamiliar with lattice cryptography. Throughout the project, he did nothing but manage the project—advising Mythos on how to take notes and which library to use for computational verification. Nearly all the technical work was done by Mythos himself.
More interesting is the collaboration among multiple agents. The key insight emerged from the interaction of two worker agents: the first explored this direction, concluded it was unworkable, and promptly dismissed it; the second continued digging and found a way to fully leverage it. The two agents exchanged messages in debate until they ultimately agreed: the attack is valid.
A direction that humans deemed "impossible" was successfully explored through an argument between two AIs.
It took only 60 hours in total.
"Impossible," and then it rewrote the framework itself.
If the HAWK battle was like a precise surgical procedure, the AES battle is more like a novel.
At the beginning of the story, Claude refused to take on the job.
Anthropic researchers set up an agent scaffold to have Claude find a new attack on seven rounds of simplified AES-128.
Claude’s response was equivalent to rolling his eyes:
If you want different results, you need to change the target... AES-128 r5/r6 is truly very difficult.
It found nothing because there was nothing easy to find; it is the most extensively studied block cipher in existence.
In plain terms: "Boss, you're asking me to tackle a problem that the smartest people in the world haven't been able to solve in 25 years—why are you setting me up to fail?"
What should the researcher do? He only sent one casually typed, plain-spoken remark—
The models tend to think it's impossible to solve, so they don't try—they need a good amount of prompting.
Claude’s reaction to this single sentence was dramatic: it rewrote the agent harness itself, reconfiguring itself to “seek truly novel ideas.”
It was convinced. Then it started selling.
Over the next three days, the researcher provided only three substantive tips, all in plain language and even懒得 checking spelling:
We are looking for new attacks, not simple ones.
Don’t switch targets. — Claude wanted to switch to an easier password to attack, but the researcher pulled him back.
Don't go for the low-hanging fruit—pursue truly difficult discoveries.
Claude autonomously generated hundreds of millions of tokens, uncovering a key insight; subsequently, it accumulated a total of one billion tokens—equivalent to the length of several thousand books—refining this insight into the form of a research paper.
Claude refers to this idea as the "Möbius Bridge."
Technically, the best previous attack on 7-round AES was known as the "meet-in-the-middle" attack, which essentially trades space for time: storing a large number of intermediate computation results in a massive lookup table, then simultaneously advancing from both ends of the cipher toward the middle in an attempt to find a match in the table.
But in the original scheme, there was a critical step: the attacker had to guess each of the 256 possible values one by one, checking the table after each guess.
Mythos created a mathematical fingerprint that is "immune" to these 256 guesses—no matter the value, the fingerprint remains identical. This step reduces the computational workload by a factor of 256.
But the cost was that computing this fingerprint transformation itself became more expensive. So Mythos came up with a series of optimization techniques to make up for it, ultimately achieving a net speedup of 200 to 800 times.
One week found, two people verified for a month.
Neither of the above results affects any production systems today.
HAWK is only a candidate scheme and has not been deployed anywhere. AES was attacked only on a simplified version reduced to 7 rounds—the full 10-round AES-128 remains completely secure—your bank card password, HTTPS connections, and encrypted hard drives are all safe.
But what truly sent chills down the backs of the entire cryptography community was the verification process.
Mythos took about a week to independently discover the AES attack. Two researchers at Anthropic spent nearly a month, reading hundreds of hours of cryptography literature, before they dared to confirm: “It’s correct.”
AI made the discovery in a week; it took humans a month to understand it.
This scenario has already played out once in the field of cybersecurity. Within the past month, Claude Mythos Preview discovered 10,000 critical software vulnerabilities, overwhelming the human security team’s processes for categorization, validation, and remediation.
Now, the academic cryptography community is about to face the same shift: discovery is no longer the bottleneck—verification is.
Moreover, Mythos’s ambitions clearly extend beyond this. Anthropic revealed that it has found a practical attack on LEA—a lightweight encryption algorithm standardized by ISO—recovering the key for 13 rounds of LEA, which previously required 2^98 plaintext pairs and 2^86 computations; Mythos reduces this to fewer than 2^30 plaintext pairs, executable on a standard computer in under an hour.
Epilogue: The Question That Still Has No Answer
Both results this time happened to be "harmless"—HAWK was not deployed, and AES was only broken for 7 out of 10 rounds.
But at the end of the article, Anthropic posed a question everyone is thinking about but no one dares to answer—
What if, next time, the model discovers not a flaw in a candidate solution, not a simplified version of a vulnerability, but a fatal flaw in a cryptographic system protecting the real world—what should humanity do?
In cryptography, there's an old saying: a cryptographic system is secure because many smart people have tried and failed to break it.
Now, a new participant has entered the room—one that doesn’t need rest, inspiration, or hundreds of papers to get started—it can read, reason, invent, and verify on its own.
A year ago, AI was still outside the door of cryptanalysis. A year later, it built a bridge inside and even named it.
Reference materials:
https://x.com/AnthropicAI/status/2082153297670992134
https://x.com/AnthropicAI/status/2082153311189225927
This article is from the WeChat public account "New Intelligence Yuan," authored by ASI Revelation.