The global cryptographic community is currently gripped by a quiet but urgent transition. For decades, the digital world has relied on RSA and Elliptic Curve Cryptography to secure everything from bank transfers to state secrets, but the looming shadow of quantum computing threatens to render these standards obsolete. This has sparked a high-stakes race to establish Post-Quantum Cryptography (PQC) standards—mathematical frameworks capable of resisting the immense processing power of a future quantum computer. In this environment, the gold standard for any new proposal is its ability to withstand rigorous cryptanalysis. Traditionally, this meant years of manual scrutiny by the world's most elite mathematicians. However, the arrival of high-reasoning AI is fundamentally altering the speed and nature of this verification process.

The Cryptanalysis Capabilities of Claude Mythos

Anthropic has recently revealed the results of an undisclosed advanced model, Claude Mythos, which has demonstrated a startling ability to identify vulnerabilities in complex cryptographic systems. The most significant achievement involves HAWK, a signature scheme that was being considered as a candidate for post-quantum standardization. HAWK is built upon the module-LIP (module-Lattice Isomorphism Problem), a mathematical foundation intended to be impervious to both classical and quantum attacks. Claude Mythos succeeded in identifying a new key recovery algorithm for HAWK, a result that effectively diminishes the likelihood of the scheme ever becoming a global standard. Notably, the model did not invent a brand-new branch of mathematics to achieve this; instead, it applied existing cryptographic tools with a level of thoroughness and precision that surpassed previous human attempts.

Beyond the realm of PQC, Claude Mythos was tasked with attacking the Advanced Encryption Standard (AES), the bedrock of modern symmetric encryption. While the model did not break the full AES implementation, it targeted a reduced-round variant consisting of seven rounds. In this controlled environment, Claude Mythos achieved a marginal performance improvement over existing research from 2013. However, the practical application of this attack remains out of reach for current hardware. To execute the attack, a malicious actor would require $2^{89}$ cryptographic operations and would need to secure $2^{105}$ encryption results for chosen plaintexts. These figures are astronomically high, placing the AES result in the category of theoretical progress rather than a practical security breach.

The methodology used to reach these conclusions is perhaps the most revealing aspect of the research. The Anthropic team did not employ domain experts to fine-tune the model or guide it through the mathematical proofs. Instead, the results were derived using prompting alone. By instructing the AI to find vulnerabilities and allowing it to iterate on its own findings, the model demonstrated a capacity to ingest vast amounts of existing cryptanalysis literature, synthesize the underlying logic, and extend those methods to create new attack vectors.

From Text Generation to Research Synthesis

This shift in capability marks a critical turning point in how the industry views large language models. For years, the primary critique of AI in specialized fields was its tendency to hallucinate or provide plausible-sounding but mathematically incorrect answers. The HAWK and AES results suggest that the intelligence ceiling is moving. The gap where a model suddenly fails—the cognitive cliff—is being pushed further outward, allowing AI to operate as a legitimate research partner in fields requiring absolute logical rigor.

The real distinction here is the difference between generative synthesis and analytical discovery. In the case of HAWK, the AI performed a full attack where the recovery of the key could be immediately verified through executable code. This creates a tight feedback loop where the AI can test a hypothesis, observe the failure or success, and refine its approach. This is fundamentally different from writing a poem or a summary; it is the application of logic to a verifiable objective. The AES result, while less practical, shows that the model can navigate the extreme nuances of theoretical cryptanalysis, even if the resulting efficiency gains are incremental.

This creates a new tension in the security world. If AI can synthesize existing literature to find vulnerabilities faster than human teams, the traditional window of safety for new cryptographic proposals is shrinking. Security designers can no longer assume that a system is safe simply because it resists known attack patterns. The ability of Claude Mythos to combine disparate tools into a novel attack means that the baseline for what constitutes a secure algorithm must be raised. We are entering an era where the primary auditor of cryptography is no longer just a human mathematician, but an AI capable of reading every paper ever published on the subject in a matter of seconds.

As these models become more capable, the industry is seeing a corresponding rise in the importance of machine-checkable proofs. Because it is increasingly difficult for humans to manually verify the minute efficiency gains an AI might claim in a theoretical attack, tools like Lean—a formal verification language—are becoming essential. The emerging workflow is a hybrid one: the AI proposes a hypothesis or a potential vulnerability, and a formal verification tool or a human expert provides the final, immutable proof. This collaboration transforms the AI from a chatbot into a high-speed hypothesis engine.

The era of relying on the obscurity of complex mathematics for security is ending, replaced by a regime of continuous, AI-driven auditing.