Claude Mythos Just Did What Human Cryptographers Couldn’t: It Broke Post‑Quantum Crypto
Anthropic has revealed that an internal, unreleased version of its most powerful AI model-nicknamed Claude Mythos-has uncovered two brand‑new cryptographic attacks. One of them targets a cutting‑edge post‑quantum digital signature scheme that is currently under consideration to become a U.S. federal standard.
In other words: an AI system, running largely on its own, managed to spot mathematical weaknesses that professional cryptographers had spent years trying and failing to find.
—
An AI vs. the Future of Encryption
The most striking result concerns HAWK, a post‑quantum digital signature scheme. Digital signatures are the mathematical backbone that proves, “yes, this transaction or message really came from me”-without ever exposing your private key.
HAWK has been designed to survive the coming era of quantum computers, which are expected to break many of today’s standard cryptographic systems. The scheme is based on lattice cryptography, a family of problems believed to be resistant even to powerful quantum algorithms.
The U.S. National Institute of Standards and Technology (NIST), the body responsible for selecting the next generation of cryptographic standards, advanced HAWK into the third round of its post‑quantum signature competition. There, it stands as the last remaining lattice‑based signature candidate in the process-exactly the kind of design expected to anchor tomorrow’s secure internet.
That’s the system Claude Mythos managed to attack.
—
Claude Spots a Hidden Symmetry in HAWK
According to Anthropic, Claude Mythos identified a previously unnoticed symmetry buried in the mathematics that underpin HAWK. No human cryptanalyst had thought to exploit it before.
By recognizing this structural pattern, the model was able to craft a novel attack-at least for the smallest, most constrained versions of the scheme’s parameters. While this doesn’t mean HAWK is instantly obsolete across all configurations, it is a serious enough finding that it could reshape how the scheme is parameterized, implemented, or even judged in the standardization process.
The key point is not only that a vulnerability exists, but that an AI was able to dig it out of an extremely abstract mathematical design that had survived years of expert human scrutiny.
—
Not Just One Break: A Second Cryptographic Attack
HAWK wasn’t the only target. Claude Mythos also developed a second, independent attack on a different cryptographic construction. Anthropic indicates this second result involved an advanced analysis related to AES, the widely deployed symmetric encryption standard used across everything from messaging apps to banking systems.
The AES‑related finding reportedly didn’t slide into place effortlessly. Engineers describe the model’s reasoning as needing “a pep talk”-a bit of prompt‑level guidance and encouragement to keep exploring promising but non‑obvious directions. Once nudged, however, Claude continued to iterate and ultimately surfaced a viable line of attack that had not been documented before.
Crucially, the discovery phase-Claude generating hypotheses, exploring structure, and sketching an attack-was substantially faster than what came next: humans and tools laboring to verify the result.
—
When Verification Takes Longer Than Discovery
Anthropic emphasizes that confirming the attacks took longer than the AI’s initial breakthroughs. In both cases, human cryptographers and external experts had to step in, painstakingly reconstructing the argument, checking the algebra, and running experiments to ensure the vulnerability was real and not merely a mirage of mistaken reasoning.
That asymmetry-fast AI‑generated ideas, slow human verification-may become a defining feature of the new era of security research. If a model like Claude can spin out dozens or hundreds of plausible attack strategies in the time it takes a person to carefully check just one, the bottleneck moves from “can we think of the idea?” to “can we vet and interpret the idea safely and correctly?”
In practical terms, these results show that large models can already act as powerful force multipliers in advanced mathematical and security domains, but only if paired with rigorous human oversight.
—
What This Means for Post‑Quantum Cryptography
The timing and target of the HAWK attack are significant. Post‑quantum cryptography is not an academic sideshow; it is a live, global effort to replace much of the encryption ecosystem before large‑scale quantum computers exist.
If NIST standardizes a system that turns out to be weak, the cost of migration, patching, and potential exploitation would be immense. Every credible attack against a candidate scheme-especially one still being evaluated-is therefore a major event.
Claude Mythos’s attack on HAWK does not automatically disqualify the scheme. Instead, it raises questions:
– Are only the smallest parameter sets affected, or does the symmetry generalize to stronger configurations?
– Can HAWK be tweaked or “hardened” to neutralize the attack while retaining efficiency?
– Should reliance on this kind of structure be reconsidered entirely for long‑term security?
Even if HAWK survives with modifications, the episode demonstrates that AI‑assisted cryptanalysis can significantly change the threat model for any proposed standard.
—
AI as a Cryptographic Red‑Team
One of the most important implications is strategic: AI models like Claude Mythos can effectively serve as automated “red teams” for cryptography.
Traditionally, a new cryptosystem is proposed, and then specialists around the world spend years trying to break it. Many schemes die quietly after someone finds an elegant attack; a select few withstand the pressure and become standards.
Now, AI can compress part of that cycle. Models can:
– Generate unusual mathematical perspectives humans might not think to try.
– Explore enormous search spaces of potential algebraic manipulations.
– Combine ideas from different branches of mathematics quickly.
– Suggest counterintuitive parameter choices that reveal edge‑case weaknesses.
None of this replaces human expertise-AI still produces dead ends, errors, and misleading lines of thought. But as a brainstorming and exploration engine, a capable model can significantly widen the net of what gets tested.
For governments, companies, and open‑source projects designing new cryptography, deliberately running AI‑based red‑teaming before deployment may soon become a best practice rather than an experiment.
—
The Double‑Edged Sword: Attackers Get AI Too
There is a darker side to these advances. If an AI can help defenders discover subtle vulnerabilities, it can also help attackers.
Today, discovering a genuinely new break in a modern cryptosystem is so difficult that only a small number of experts have the skills, patience, and background to try. If powerful models lower that barrier-even slightly-the odds increase that non‑specialists or small groups could unearth dangerous flaws.
The current HAWK result offers a glimpse of a future where:
– State‑level actors use AI systematically to trawl through public algorithms for exploitable structure.
– Criminal organizations experiment with models to identify weak implementations, poorly chosen parameters, or legacy systems ripe for attack.
– Aging cryptographic standards fall not because humans finally get around to scrutinizing them, but because AI goes hunting.
This reality adds urgency to proactive, AI‑assisted defense. The same tools that might supercharge offensive cryptanalysis must be harnessed early and aggressively by the defenders responsible for securing infrastructure.
—
Why the HAWK Break Matters to Crypto and Finance
For the digital asset ecosystem, post‑quantum signatures are not a remote academic topic. Blockchains, wallets, cold storage systems, and institutional custody solutions all rely heavily on digital signatures. If quantum computers eventually make today’s signatures obsolete, every serious project will be forced to migrate.
Systems like HAWK are candidates for that future. An AI‑discovered weakness, even in limited parameter regimes, sends a clear signal to the industry:
– Relying on “unbroken so far” as a security argument is weaker in an AI era.
– Future‑proofing wallets and protocols will require cryptosystems that have survived both human and AI scrutiny.
– Formal verification, rigorous reduction proofs, and conservative assumptions will matter more than ever.
While no major blockchain is using HAWK today, the episode is a real‑world test case in how quickly a seemingly solid post‑quantum scheme can face new questions once AI joins the review process.
—
How Did Claude Mythos Actually Do It?
Anthropic describes Claude Mythos as a locked, internal model-more powerful than the public Claude releases-tuned for deep reasoning tasks in mathematics, programming, and formal analysis.
The company reports that the model operated roughly as follows:
1. Understanding the Scheme
It first digested the formal specification of the cryptographic algorithms: definitions, parameter sets, and underlying hard problems.
2. Exploring Structure
Claude then probed the algebraic and geometric structure of the constructions, searching for symmetries, invariants, and relationships between components that humans might regard as “just implementation details.”
3. Hypothesizing Attacks
When it noticed a potentially exploitable symmetry in HAWK, the model sketched attack strategies-outlining how certain inputs or parameter ranges might make the scheme behave in a less secure way.
4. Refining With Guidance
Human researchers guided the model away from obviously wrong tracks and encouraged it to flesh out promising angles, especially for the AES‑related work, where its reasoning needed to be kept on target.
5. Hand‑Off for Verification
Once a potential break was outlined, humans took over to formalize, test, and validate the ideas against real cryptographic instances.
This workflow underscores a crucial pattern: AI as idea generator, human as validator and system‑level decision‑maker.
—
A New Standard for “Secure Enough”?
Cryptographers traditionally lean on conservative assumptions: a scheme is considered secure if no efficient attack is known, despite robust efforts from experts. What happens when highly capable AI becomes part of that expert set?
In the near future, “we tried to break it and failed” may no longer be satisfying without specifying:
– Which AI models were used, and how capable they were.
– How long they were run, under what prompts and constraints.
– Whether they were allowed to search through non‑obvious structural transformations or just check standard lines of attack.
Over time, the bar for calling something “secure” could rise to include comprehensive, documented AI‑assisted red‑teaming. A system that survives years of human scrutiny but falls quickly to a few days of AI exploration may be viewed as fundamentally less trustworthy.
—
Preparing for an AI‑Intensive Security Landscape
For organizations designing or implementing cryptography-whether for financial systems, encrypted messaging, or critical infrastructure-the Claude Mythos results highlight several practical steps:
– Assume AI‑assisted adversaries: Treat any widely known algorithm as if it will eventually be probed by powerful models looking for weak points.
– Integrate AI into your own review: Use capable models under expert supervision to stress‑test new designs and configurations.
– Favor battle‑tested, conservative choices: Novelty in cryptography is exciting, but in an AI era, conservative parameter selections and well‑understood constructions may offer safer ground.
– Plan for agility: Systems should be architected to support cryptographic agility-the ability to migrate to new schemes if standards shift or vulnerabilities emerge.
The big lesson is not that cryptography is doomed, but that the tempo of discovery has changed.
—
The Road Ahead
Claude Mythos breaking a post‑quantum signature candidate and devising a second cryptographic attack is a watershed moment. It shows that frontier AI models can do more than summarize papers or write code-they can contribute original, high‑impact insights in fields previously considered the domain of a small elite.
For defenders, this is an opportunity: cryptographic designs can be hardened earlier and more thoroughly. For standards bodies, it is a warning: no evaluation process can ignore AI‑based analysis and still claim to be exhaustive.
And for anyone who relies on cryptography-which, in the digital economy, is effectively everyone-the message is clear: the definition of “secure by design” is evolving. In the post‑quantum, post‑human‑only era of cryptanalysis, AI like Claude Mythos will be part of both the problem and the solution.