Anthropic has used its Claude Mythos system to uncover previously unknown weaknesses in a leading post quantum signature candidate and in a reduced version of AES, signaling a new phase in AI assisted cryptanalysis. These results matter because they show modern AI systems can now participate in serious mathematical security work, potentially reshaping how encryption standards are designed, reviewed and trusted.
Why this moment in cryptography and AI matters
For most people and most businesses, encryption is invisible until something goes wrong, yet it underpins online banking, messaging, cloud services and almost every digital transaction. The algorithms at the heart of that encryption are carefully designed and vetted by expert cryptographers, often over many years, because a single mistake can expose data at global scale.
The announcement that Claude Mythos Preview helped uncover weaknesses in a post quantum digital signature scheme called HAWK and in a seven round research variant of AES 128 is therefore notable on two levels. First, it touches on NISTs ongoing effort to standardize next generation encryption that can resist future quantum computers. Second, it shows an AI system performing original mathematical cryptanalysis rather than merely automating known techniques, which is a step beyond familiar uses of AI in code generation or bug finding.
Claude Mythos exposes research weaknesses in HAWK and 7‑round AES, marking AI’s entry into serious post‑quantum cryptanalysis and NIST‑grade security review
This is not a story about online bank transactions suddenly becoming unsafe. Both attacks target research configurations, not the parameters currently used to protect real world systems, and NIST and independent analysts emphasize that deployed AES and approved post quantum schemes remain secure given current evidence. The significance lies in what this reveals about the future of cryptographic research and standardization when powerful AI systems are part of the toolkit. Crucially, these results mean there are no production systems at risk from the discovered weaknesses even as they raise expectations for how AI will be integrated into future cryptographic evaluations.
Background: how HAWK and AES fit into the security landscape
HAWK is a lattice based hash and sign digital signature scheme that NIST advanced to later rounds in its additional post quantum signature process. It is related to earlier lattice based designs such as Falcon and was attractive because it aimed to deliver strong security with relatively efficient signatures and keys compared to some competitors.
NISTs process for post quantum cryptography has been underway for years, with multiple rounds of public scrutiny, formal proofs and practical attacks on candidate schemes. In that context HAWK appears as the only lattice based scheme among nine signature candidates that NIST moved into a third round for additional evaluation in twenty twenty six. The team behind HAWK even provided a smaller challenge parameter called HAWK 256 specifically as a target for cryptanalysts to stress test the design.
AES, in contrast, is already a global workhorse. The standard AES 128 cipher uses ten rounds of substitution and permutation to scramble data and is widely deployed in protocols that protect web traffic, storage systems and financial data. Cryptographers often study reduced round versions such as seven round AES 128 under controlled conditions to understand safety margins and the gap between theoretical attacks and practical exploitation.
Claude Mythos focused precisely on these research settings rather than on deployed configurations, which is why the results are interesting for experts yet reassuring for operators of current systems.
What Claude Mythos actually discovered
The HAWK key recovery attack
Working with Mythos Preview, an Anthropic researcher developed an end to end key recovery attack against the HAWK 256 parameter set. HAWK 256 was intended as a challenge instance rather than a production security level, but it still represented a strong target in the post quantum design space.
The attack exploits a previously unused symmetry in the lattice structure underlying HAWK, building what the researchers describe as a tau cocycle lattice from the public key and then using lattice reduction and sieving to recover short vectors that help reconstruct a secret signing basis. In practical terms, this reduces the effective key strength by roughly a factor of two, meaning that a brute force style attack becomes substantially easier than originally believed.
For the smallest HAWK key size, the estimated effort to recover a private key falls from around two to the sixty fourth power operations to about two to the thirty eighth power operations. That is a reduction of about sixty seven million times in the amount of work needed. Importantly, this improvement came after roughly two years of human cryptanalytic review had failed to uncover an attack of comparable strength, while Mythos arrived at the strategy in about sixty hours of focused experimentation on large computing resources.
NIST and Anthropic stress that the attack remains exponential time and is not currently practical for the larger HAWK parameter sets that correspond to proposed production security levels. HAWK 512 and HAWK 1024 are still considered secure under known attacks, and HAWK 256 was explicitly framed as a research challenge rather than a recommended deployment choice. However, the fact that key strength must effectively be doubled to restore the intended security level undermines some of the original efficiency advantages that made HAWK appealing.
The Möbius Bridge attack on seven round AES 128
The second major finding concerns a seven round research variant of AES 128, not the full ten round cipher used in real systems. Previous best known attacks on seven round AES relied on meet in the middle techniques that required checking a large set of intermediate values, including a costly enumeration of two hundred fifty six possible byte guesses.
Claude Mythos produced a new mathematical fingerprinting method called Möbius Bridge that remains invariant across all of those guesses, eliminating the need to test each one individually. By removing this guessing step, the attack achieves an estimated speedup between two hundred and eight hundred times compared with earlier approaches for the same configuration. This does not break full AES 128 and still demands substantial amounts of chosen plaintext data, but it tightens the known boundary between safe and unsafe numbers of rounds for this widely used cipher.
From a cryptographic engineering perspective, these results sharpen the understanding of how many rounds can be peeled away in theoretical analyses before an attack becomes realistically feasible in a high resource setting. This kind of insight is valuable when designers consider lightweight variants, protocol compromises or special purpose hardware implementations that might be tempted to trim margins.
How strong and how worrying are these attacks
The first question that security teams and non specialists ask is whether this means existing encryption is broken. The short answer is no for the configurations that protect everyday traffic and data.
The HAWK attack currently applies to the HAWK 256 challenge parameter and serves as a wake up call rather than an immediate operational threat. NIST has confirmed the findings and noted that the stronger HAWK parameters remain secure against this specific style of attack, which is tailored very closely to the structure of HAWK and does not generalize to other lattice based schemes. Nevertheless, having to double key sizes to compensate for the discovered weakness will likely affect how HAWK is evaluated in the remaining rounds of the NIST process.
The Möbius Bridge attack on seven round AES is more of a conceptual advance than a direct route to real world exploitation. It requires very large amounts of controlled data, targets a reduced round cipher and does not touch the ten round AES variants used in protocols that safeguard most internet and financial traffic. Cryptographers view it as an important refinement of theoretical knowledge about AES rather than a break that would prompt emergency patching.
Where the concern does grow is in the demonstration that a frontier AI system can autonomously explore and exploit deep mathematical structures at scale, in ways that even dedicated human teams had not found. Mythos was given substantial computational budget and ran for around a week of automated research, producing work that would traditionally have required a specialist cryptographic lab and months or years of concentrated effort.
This shifts the perceived balance between offense and defense in cryptography. Well resourced actors who can afford large AI runs now have a stronger tool for finding weaknesses, but so do standards bodies and responsible companies that can use similar methods to harden designs before deployment.
Historical context: AI moving from pattern recognition to mathematical discovery
The direction of travel in AI research has been toward systems that go beyond pattern recognition in images or text and begin to tackle formal mathematics, theorem proving and algorithm discovery. Previous work has shown AI models helping to find new matrix multiplication algorithms, assisting with theorem proving in formal systems and suggesting novel scientific hypotheses, but those projects were largely confined to academic settings and carefully curated tasks.
Claude Mythos brings that trajectory directly into applied security research. It is deployed as a frontier reasoning system with access to formal specification, code and cryptographic constructions, then guided by human researchers to search for non obvious algebraic structures and exploitable symmetries. The HAWK attack is a concrete example of AI locating a nontrivial automorphism in a lattice where theorists suspected that such a structure might exist but had not been able to reach it in practice.
This interplay between human intuition and automated exploration echoes earlier advances in both mathematics and computer science, where tools such as computer algebra systems or SAT solvers amplified expert capability rather than replacing it. The difference now is scale and generality. Large AI models can generate and test thousands of candidate attacks, transformations and invariants in parallel, guided by probabilistic assessments of what might be promising, and then refine those ideas with human feedback.
For cryptography, which depends on systematically exhausting attack avenues rather than relying solely on proofs, this is a profound change in the available toolkit.
Implications for technology, businesses and society
For technology teams the practical implication is that AI assisted cryptanalysis is now something that must be taken seriously in threat modeling and standardization discussions. Cloud providers, big financial institutions and governments already rely on strong cryptographic standards, and they participate in processes like NISTs precisely to avoid surprises.
The ability of AI systems to uncover substantive weaknesses in candidate algorithms before those algorithms are widely deployed is ultimately a positive development for defense, but it also raises the bar for what counts as a thorough review.
Businesses that depend on long term cryptographic assurances, such as banks or health care providers, can take comfort in the fact that NIST and other bodies remain cautious and responsive when new information appears. The HAWK episode illustrates how a candidate can be stress tested, weaknesses publicized and parameters adjusted or even designs reconsidered before standards are finalized, all without causing disruption to existing systems.
At the same time, there is an arms race dimension. If a commercial AI model can achieve this level of cryptanalytic performance with published techniques and significant but not unimaginable computing budgets, then state actors and large corporations will also explore similar capabilities. This heightens the importance of transparency, responsible disclosure and governance around powerful AI models that can be directed toward offensive security research.
Societally, the episode contributes to a broader narrative about AI as both tool and potential hazard. On the one hand, AI can help find latent vulnerabilities in algorithms that would otherwise be standardised and deployed for decades, thereby preventing future disastrous breaks. On the other, systems with this level of capability could in principle be misused to probe legacy protocols or less well studied cryptographic constructions that do protect live systems, particularly if combined with expertise and tailored hardware.
Public trust in encryption depends on confidence that the strongest known attacks have been explored and addressed. The more AI is involved in that exploration, the more important it becomes to maintain rigorous peer review, open publication of technical results and clear communication about what has and has not actually been broken.
Comparing with earlier developments in AI and security
Earlier uses of AI in security have focused on anomaly detection, malware classification, vulnerability scanning and code assistance, often improving speed or coverage but not fundamentally changing the mathematics of cryptographic primitives. Machine learning models have been very effective at finding bugs in complex software or at spotting unusual network traffic, yet the core encryption algorithms remained evaluated primarily by human mathematicians and specialized tools.
Claude Mythos represents a deeper integration between AI and the hardest parts of security research. It is not simply ranking known vulnerability patterns but helping to generate entirely new lines of attack on advanced schemes that had already survived rounds of expert scrutiny. That level of contribution narrows the gap between experimental AI mathematical reasoning and practical security outcomes.
In that sense, this work sits alongside early examples of AI discovering non obvious results in formal domains as a signal that future standards processes will likely expect AI assisted exploration as part of their due diligence. It would not be surprising to see future NIST style calls for proposals explicitly reference AI based cryptanalysis as an expected element of candidate evaluation and community feedback.
Opportunities and risks in AI assisted cryptanalysis
The opportunity side is clear. If powerful AI systems can regularly uncover weaknesses in candidates like HAWK before those algorithms are enshrined in standards, the resulting cryptographic ecosystem will be more robust. Designers can use AI to test structural assumptions, search for symmetries, and explore parameter spaces that are too large for manual analysis, then patch or discard designs that fail those tests.
AI can also help teach the next generation of cryptographers by providing interactive environments where students explore attacks, understand algebraic constructions and see how subtle design choices affect security, backed by automated reasoning and simulation. This may widen the pool of experts who can work on these problems.
The risks are equally real. Access to high powered AI cryptanalysis may become uneven, with well funded actors gaining a strategic advantage over smaller organizations and open academic groups. If proprietary models and private research uncover weaknesses in deployed systems but those findings are not disclosed promptly and responsibly, the mismatch between offensive and defensive knowledge could be exploited.
There is also a governance challenge around dual use. Models like Claude Mythos can be used to stress test candidate algorithms in cooperation with standards bodies, but they could also be directed toward legacy systems, embedded devices or niche protocols where the impact of a break would be immediate. Managing access, monitoring use and embedding safeguards without stifling beneficial research will be a central question for regulators and AI providers.
Key takeaways and what to watch next
Several clear lessons stand out from the HAWK and AES findings.
First, frontier AI systems are now capable of contributing meaningfully to modern cryptanalysis and mathematical security research, not just automating routine tasks. That capability can and should be harnessed by standards bodies and responsible organizations to harden future algorithms before deployment.
Second, the specific weaknesses discovered do not endanger current mainstream encryption, but they do reshape the evaluation of at least one post quantum candidate and refine understanding of AES security margins. This is exactly the kind of course correction that robust processes are designed to accommodate.
Third, the broader security and governance community needs to treat AI assisted cryptanalysis as a real capability with both defensive and offensive potential, and to develop norms, controls and collaborations that keep its use aligned with public interest.
Watching how NIST responds to the HAWK results, how other candidate schemes fare under similar AI assisted scrutiny, and whether new frameworks for responsible AI use in security emerge over the next few years will give a good sense of how quickly this new phase of cryptographic research is maturing. The underlying message is that encryption will not stand still in the age of AI, and that the most trustworthy systems will be those that embrace rigorous, transparent, AI informed testing as part of their lifecycle.
Conclusion
Claude’s new cryptographic breakthroughs mark a quiet but genuine turning point for digital security. An advanced version of the model, Claude Mythos Preview, has independently discovered real mathematical weaknesses in two modern cryptographic schemes, demonstrating that large scale AI systems are now capable not just of helping with security work but of doing original codebreaking research in their own right.
Why this matters now
For most people, cryptography is invisible, yet it underpins nearly every online interaction from banking and messaging to cloud storage and government systems. The long standing assumption has been that well vetted algorithms, once standardized and broadly deployed, offer strong protection until a major theoretical breakthrough or a dramatic change in computing power arrives.
Claude’s results do not break the cryptography that protects logins or payments today, but they do show that AI can discover subtle weaknesses in candidate algorithms and testbed variants much faster than traditional human driven teams. That shift matters for governments planning post quantum standards, for companies that rely on cryptographic libraries, and for defenders who have assumed that algorithm level breakthroughs would remain rare and slow.
Perplexity Sonar’s analysis of these events has emphasized precisely this inflection point. The story is less about an immediate security crisis and more about a change in who can do advanced cryptanalytic research and how quickly that research can iterate.
How we got here: AI meets cryptography
Cryptography and AI have been moving toward each other for years. Security researchers have used machine learning to analyze side channel signals, detect implementation bugs, and automate parts of vulnerability discovery. At the same time, adversarial use of AI has grown, with models that can sift vast telemetry, identify misconfigurations, and optimize attack paths far beyond what manual operators can manage.
Recent work points to a deeper structural change. Studies on AI driven adversaries argue that practical cryptographic failure often occurs not because the underlying mathematics is broken, but because real world implementations leak information or expose keys in ways that smart models can exploit. Other research tracks how AI helps both attackers and defenders accelerate post quantum migration, rewrite legacy code that still depends on RSA or ECC, and scan implementations for subtle side channel flaws.
Claude Mythos Preview fits directly into this trajectory. Earlier experiments showed the model could autonomously identify and exploit vulnerabilities across major cryptographic libraries and protocols including TLS, AES GCM, and SSH by finding mistakes in how the algorithms were implemented in code. That work already hinted that modern AI systems could act as powerful automated security reviewers. The latest research moves a step closer to classical cryptanalysis by uncovering weaknesses in the mathematical structure of algorithms themselves.
What exactly did Claude find
Anthropic’s team focused Mythos Preview on two targets that matter for the future of secure communications.
HAWK digital signature scheme
HAWK is a proposed digital signature algorithm designed for a post quantum world, intended to remain secure even against powerful quantum computers. It has been under expert scrutiny for roughly two years and is one of several schemes being evaluated by standards bodies such as the United States National Institute of Standards and Technology for eventual adoption.
In about sixty hours of focused automated work, Claude Mythos Preview discovered a previously unknown attack that effectively cuts HAWK’s security level in half. The attack does not completely break the scheme but makes it significantly less robust than designers intended, forcing a reevaluation of its parameters and of its candidacy as a long term standard. Importantly, HAWK has not been deployed in production systems, so no live infrastructure is currently exposed.
Round reduced AES cipher
The second result concerns a weakened version of the Advanced Encryption Standard, AES, which is the most widely used symmetric cipher protecting web traffic, Wi Fi, data at rest, and many other core systems. Researchers often study round reduced variants of AES that use fewer encryption rounds than the standard ten, as a way to test analytic methods without touching production systems.
Claude Mythos Preview found a new way to attack such a reduced round AES variant that speeds up known cryptanalytic techniques by a factor between two hundred and eight hundred. This is not a break of full strength AES as deployed in browsers, banking apps, or wireless routers today. However, the speedup is scientifically significant because it shows that a general purpose AI system can improve on expert designed attacks against widely studied ciphers, and can do so without being hand coded with cryptanalytic heuristics.
No immediate crisis, but a real shift in capability
A key point emphasized by both Anthropic and independent coverage is that these discoveries do not threaten current encryption systems. HAWK is not yet in production, and the AES variant under study is deliberately weakened for research purposes.
This matters for public trust. It means people’s passwords, banking transactions, and private messages are not suddenly at risk from Claude’s specific attacks. It also demonstrates responsible disclosure and careful scoping by the research team, who chose targets and configurations that would illuminate the frontier of cryptanalysis without undermining real world protections.
The deeper shift is about capability rather than immediate impact. Claude’s performance indicates that large models can autonomously explore vast search spaces of potential attacks, evaluate patterns, and converge on effective strategies faster than many human teams could. In other domains, similar models have shown the ability to orchestrate complex tasks, reason over code bases, and adapt strategies over time. When that kind of agentic behavior is turned toward cryptography, the result is a new class of AI tools that can systematically probe both algorithms and implementations for weaknesses.
Implications for technology and business
The most direct implication is for standards bodies and cryptographic researchers. Every candidate algorithm for post quantum adoption now has to be evaluated not only against human analysts and traditional techniques, but also against the capabilities of large AI systems that can run continuous, high volume search for weaknesses. Work that once depended on a relatively small community of experts may now be accelerated and scaled, potentially surfacing issues earlier in the standardization process.
For businesses, especially those in finance, cloud computing, and critical infrastructure, the lesson is more strategic. Cryptography has often been treated as a stable foundational layer, something that can be set once and revisited only when a major new standard emerges. Claude’s research suggests that the review cycle for algorithms and implementations may need to become more dynamic. Companies will likely need continuous testing of their cryptographic stacks with AI assisted tools that look for both mathematical and implementation level weaknesses.
Board members and chief information security officers are already being advised to account for AI and quantum risks in their cyber strategies, recognizing that advances in hardware and algorithms are accelerating the timeline for practical cryptanalysis. The ability of AI systems to triage large volumes of ciphertext, prioritize high value targets, and integrate with emerging quantum decryption techniques raises the stakes for long lived data such as archived diplomatic cables, health records, and industrial secrets. Even if today’s algorithms remain secure, stored encrypted data can become vulnerable once more capable attacks are feasible.
Societal and ethical dimensions
On the societal side, Claude’s discoveries highlight a dual use reality. The same AI techniques that help find and fix cryptographic weaknesses can, in other hands, be used to exploit those weaknesses and undermine privacy at scale. Research on AI driven adversaries shows how models can optimize attacks against implementation level observability, targeting side channels, entropy failures, and key lifecycle mistakes rather than brute forcing algorithms directly.
This raises ethical and governance questions. Who should be allowed to run agentic AI systems against cryptographic infrastructure, and under what oversight. How should results be disclosed so that defenders can patch issues without giving attackers a detailed road map. What safeguards should be embedded into AI platforms to prevent unrestricted use for offensive cryptanalysis.
Perplexity Sonar’s treatment of the Anthropic work reflects a growing consensus that transparency and context are crucial. Simply announcing that an AI model has found cryptographic flaws is not enough. Responsible reporting demands clear explanations of what is and is not at risk, how experiments were scoped, and what mitigations are available to practitioners.
Comparing with earlier developments in AI security
There have been notable earlier moments where AI has impacted security, such as using machine learning to improve intrusion detection, automate threat hunting, or optimize side channel analysis. In each case, AI acted as an accelerator for existing techniques, making familiar tasks faster or more scalable.
Claude’s cryptographic work differs in two ways. First, it moves closer to core mathematical weakness discovery, rather than focusing only on noisy signals or implementation bugs. Second, it shows an autonomous system exploring a sophisticated problem space and generating novel attacks without being explicitly programmed with the underlying theory. This resembles AI assisted scientific discovery more than traditional security tooling.
Earlier AI contributions to post quantum migration focused on rewriting code, scanning for hard coded primitives, and helping enterprises inventory where legacy algorithms were used. The new results suggest that future migration plans will need a feedback loop that includes AI systems testing candidate algorithms, scrutinizing their implementations, and flagging schemes that prove fragile under automated cryptanalysis.
What defenders should do next
For practitioners, the practical response to Claude’s discoveries is twofold.
First, harden cryptographic algorithms and implementations with AI in mind. This means adding automated testing pipelines that use large models to probe libraries, protocols, and hardware for weaknesses, treating AI as an integral part of assurance rather than an optional experiment. It also means staying engaged with evolving standards and research, since schemes considered safe today may be reevaluated once AI assisted attacks become more common.
Second, build guardrails around advanced AI systems capable of cryptanalysis. Organizations that develop or deploy such models should establish clear policies about permissible use, logging and oversight, and responsible disclosure pathways when new weaknesses are found. Guardrails are not meant to blunt scientific progress; they are meant to ensure that discoveries lead to improved defenses rather than uncoordinated exploitation.
Security has always been a moving frontier. Cryptographic history is full of once trusted algorithms, from early block ciphers to now deprecated hash functions, that were eventually superseded when new attacks emerged. The difference now is speed and scale. AI systems like Claude can compress years of manual analytic work into hours or days, shifting the balance between attackers and defenders.
Claude’s discoveries leave the field at an uneasy inflection point, exposing the limits of long trusted schemes while confirming that AI will be embedded in future cryptanalysis. Researchers now face a dual obligation. They must harden algorithms before hostile actors exploit similar capabilities, and they must build guardrails around systems that can probe the world’s defensive fabric. The episode closes not with panic, but with recognition that security’s frontier has shifted and that vigilance must evolve accordingly for decades ahead. reddit
1 comment
Comments are closed.