Artificial intelligence safety and research firm Anthropic has released two significant cryptanalysis papers generated by its unreleased, advanced AI model, internally designated as Claude Mythos. The findings, published alongside a detailed technical blog post, demonstrate the model’s expanding capacity to synthesize existing cryptographic research and formulate novel attacks against both post-quantum signature schemes and traditional symmetric ciphers. While the developments highlight the rapid evolution of machine learning in specialized mathematical domains, they also underscore critical challenges regarding the verifiability of AI-driven computational proofs.
The research focuses on two distinct areas of cryptography: a successful key-recovery attack against HAWK, a non-standard post-quantum signature scheme based on the module Lattice Isomorphism Problem (module-LIP), and an incremental improvement on cryptanalytic attacks against a reduced-round variant of the Advanced Encryption Standard (AES).
Main Facts and Technical Scope
The first major result targets HAWK, a lattice-based signature scheme designed to remain secure against attacks launched by quantum computers. Lattice-based cryptography relies on the geometric properties of high-dimensional lattices, which are believed to be difficult for both classical and quantum algorithms to solve. Anthropic’s model successfully devised a new key-recovery algorithm against HAWK. Domain experts analyzing the code note that the breakthrough did not rely on exotic mathematical shortcuts or unprecedented theoretical frameworks; rather, it stemmed from a remarkably thorough and rigorous application of existing cryptanalytic tools. By systematically executing known techniques at a scale and speed difficult for human researchers, Claude Mythos identified vulnerabilities that may effectively disqualify HAWK from ongoing post-quantum standardization processes.
The second finding involves a cryptanalytic attack on a 7-round variant of AES. AES is a foundational symmetric block cipher standardized by the National Institute of Standards and Technology (NIST) in 2001, featuring standard key sizes of 10, 12, or 14 rounds. Because breaking a full, deployed cipher remains computationally intractable, cryptanalysts routinely evaluate weakened, reduced-round versions of algorithms to gauge their structural integrity.
While headlines regarding an "attack on AES" frequently incite concern, the Anthropic result represents an incremental optimization rather than a foundational breach. The newly formulated attack requires $2^89$ cipher operations and presupposes that an attacker has already induced a target encryptor to generate $2^105$ encryptions of chosen plaintexts under a single secret key. These operational parameters remain entirely impractical in real-world deployment scenarios. Furthermore, because the attack cannot be physically executed in its entirety, the reported speedup functions primarily as an on-paper theoretical analysis, offering a modest constant-factor improvement over similar research published in 2013 by cryptographic teams at the École Normale Supérieure and related academic institutions.
Chronology and Development Methodology
The release of these findings marks a notable shift in how advanced machine learning models interact with academic research. According to Anthropic’s accompanying documentation, the research team did not construct a specialized, deeply tuned ensemble of domain-specific cryptographic experts. Instead, the process relied on generalized prompt engineering, directing the model to investigate specific algebraic structures and iterate upon known methodologies until viable attack vectors emerged.
The timeline of AI-assisted mathematics has accelerated significantly over the preceding twenty-four months. Early models demonstrated proficiency in basic coding, theorem-proving assistance, and summarizing academic literature. The emergence of systems capable of generating novel cryptanalytic pathways indicates that frontier models have crossed a threshold into autonomous synthesis. By digesting decades of published cryptographic literature—ranging from side-channel analyses to algebraic cryptanalysis—models like Claude Mythos can connect disparate concepts across subfields of mathematics.
Supporting Data and Comparative Analysis
To contextualize the HAWK and AES findings, security analysts draw comparisons to historical evaluations of cryptographic standards.
-
Post-Quantum Standardization: The National Institute of Standards and Technology has spent over a decade evaluating post-quantum cryptography (PQC) algorithms to replace legacy public-key systems like RSA and Elliptic Curve Cryptography (ECC), which are vulnerable to Shor’s algorithm on a sufficiently large quantum computer. Schemes based on lattices, such as CRYSTALS-Kyber and CRYSTALS-Dilithium, were selected for standardization in 2022. Alternative proposals like HAWK were evaluated for their specific efficiency and signature size trade-offs. The discovery that standard cryptanalytic toolkits, when applied via automated reasoning, can compromise HAWK underscores the fragility of newer, less-vetted mathematical hardness assumptions.
-
Symmetric Resilience: Symmetric ciphers like AES derive their security from the diffusion and confusion properties built into substitution-permutation networks. Decades of intensive cryptanalysis—including confidential evaluations performed by agencies such as the National Security Agency (NSA)—have consistently demonstrated that full-round AES possesses a substantial security margin. Anthropic’s work on 7-round AES confirms that while automated models can optimize complexity bounds, they have not yet discovered fundamental flaws in the core design principles of modern symmetric encryption.
The Verification Bottleneck
As artificial intelligence systems generate increasingly complex mathematical artifacts, the primary bottleneck in scientific research has shifted from generation to verification. Independent researchers note that while models are highly efficient at producing outputs that appear mathematically rigorous, they frequently generate subtle errors, logical inconsistencies, or misapplied theorems that require meticulous human review.
For attacks that run in constrained environments—such as the HAWK key-recovery script, which executes in a matter of hours against a reduced parameter set—empirical verification is straightforward: researchers can run the code and confirm whether it successfully extracts keys. However, for theoretical speedups and high-complexity analyses like the AES reduced-round evaluation, empirical testing is impossible due to resource constraints.
In response, the mathematical community has increasingly relied on formally verified proof assistants, such as the Lean programming language, to check the validity of machine-generated proofs. Yet, even formal verification systems remain sensitive to how theorem statements are initially formulated, necessitating continuous oversight by human domain experts to ensure that the proven theorem accurately reflects the real-world cryptographic problem.
Broader Industry Implications and Outlook
The publication of these results carries distinct implications across multiple sectors, ranging from enterprise security to academic research and AI governance.
Implications for Cryptographic Users
Organizations currently executing the multi-year migration to post-quantum cryptography must carefully evaluate the selection of standardized algorithms. The vulnerability exposed in HAWK demonstrates that emerging public-key schemes outside the core NIST-standardized suite face heightened risks from automated cryptanalysis. Conversely, users of symmetric cryptography (such as AES-256 for data-at-rest encryption) face minimal immediate threat. Because symmetric algorithms are constructed with deep structural safety margins, the introduction of automated intelligence-hours has not bridged the gap required to compromise full-round ciphers.
Implications for the Scientific Community
For cryptographers and mathematicians, the integration of advanced AI models represents both a powerful collaborative tool and a disruptive force. Researchers now possess computational assistants capable of accelerating literature reviews, testing hypotheses, and optimizing algebraic reductions at unprecedented speeds. However, this acceleration threatens to overwhelm traditional peer-review pipelines. The scientific community faces open questions regarding attribution for machine-assisted discoveries and the establishment of robust, scalable standards for verifying computer-generated cryptanalytic proofs.
Implications for AI Policy and Capability Scaling
The demonstration that general-purpose frontier models can conduct autonomous cryptanalysis challenges minimalist interpretations of AI capabilities. Rather than serving merely as advanced autocomplete engines, models exhibiting this level of specialized synthesis demonstrate genuine progress in complex problem-solving. This capability reinforces ongoing policy discussions regarding the oversight, deployment restrictions, and safety evaluations required for unreleased, high-end AI architectures.
As academic and commercial laboratories continue to refine these models, the boundary between human-led and machine-assisted research will continue to blur. The security of global digital infrastructure will increasingly depend on the ability of researchers to harness these same computational tools to proactively audit cryptographic systems before malicious actors can deploy them at scale.
