Anthropic's Claude AI Discovers Novel Weaknesses in Post-Quantum Cryptography Candidates

Anthropic's Claude Mythos Preview has achieved a major milestone in artificial intelligence and cybersecurity: autonomously discovering novel mathematical weaknesses in widely studied cryptographic algorithms. The breakthrough demonstrates that frontier AI models are crossing a threshold from simply assisting human coders to conducting original, expert-level mathematical research.[1][5]

The frontier AI model identified flaws in two distinct encryption methods, including a candidate algorithm designed specifically to withstand attacks from future quantum computers. In both cases, the AI uncovered mathematical shortcuts that had eluded human cryptographers during years of rigorous peer review.[2][4]

While the word "vulnerability" often triggers alarm in the cybersecurity world, security experts emphasize that these discoveries do not affect any systems currently in production. Instead, the findings represent a triumph for defensive research, demonstrating how AI can stress-test our digital infrastructure before it is ever deployed in the real world.[1][7]

The most consequential discovery involves HAWK, a digital signature scheme that was under evaluation by the U.S. National Institute of Standards and Technology (NIST). Digital signatures are the mathematical mechanisms that verify authenticity whenever a user sends cryptocurrency, logs into online banking, or accesses a secure website.[2][6]

NIST has been running a multi-year competition to identify "post-quantum cryptography" (PQC) standards. These are algorithms capable of securing data against the eventual arrival of powerful quantum computers, which could theoretically break the encryption methods currently protecting the internet.[3][6]

HAWK had survived two full rounds of expert review over two years without any significant mathematical flaws being surfaced. It was considered a strong candidate due to its compact efficiency, making it highly attractive for widespread standardization.[4][6]

Operating in a semi-autonomous agentic harness, Claude Mythos Preview spent roughly 60 hours analyzing HAWK's underlying mathematics. The model conducted an extensive literature review, performed complex mathematical reasoning, and built an end-to-end verification pipeline to confirm its findings.[1][4]

The model ultimately discovered a "nontrivial automorphism"—a hidden mathematical symmetry within the complex lattice structure that HAWK relies on for its security. Prior research had established that finding such a symmetry would enable an attack, but human experts had not determined whether one actually existed in HAWK's specific lattice.[2][6]

Exploiting this newly discovered symmetry effectively cuts the algorithm's key strength in half. It reduces the estimated computational work required to attack HAWK's smallest challenge parameter from roughly 2^64 operations down to just 2^38 operations.[6]

To restore HAWK's intended security level, its developers would need to double the size of its cryptographic keys. However, doing so would erase the compact efficiency that made the algorithm a competitive candidate in the first place.[2][3]

Following standard responsible disclosure protocols, Anthropic notified HAWK's designers in June. Acknowledging that the required fixes would make the algorithm uncompetitive, the developers officially withdrew HAWK from the NIST competition on July 29.[2][3]

In a separate experiment, Claude Mythos Preview was tasked with analyzing the Advanced Encryption Standard (AES), the ubiquitous cipher adopted by NIST in 2001 that protects everything from HTTPS web traffic to classified government data.[1][4]

The AI did not break the full 10-round AES standard used in the real world. Instead, it analyzed a simplified, 7-round version commonly used by researchers to measure security margins and understand theoretical limits.[2][7]

Working largely on its own, the model autonomously invented a novel mathematical shortcut it named the "Möbius Bridge." This new technique accelerated the best-known attacks on the reduced AES variant by 200 to 800 times.[2][4]

Even with this massive speedup, the attack remains purely theoretical. Executing it would require an impossible amount of target data—over 400 octillion messages—ensuring that real-world AES encryption remains completely safe.[2]

The true significance of the AES result lies not in immediate danger, but in the methodology. The fact that an AI model was able to invent and name a mathematical technique that had eluded human cryptographers for decades marks a paradigm shift in security research.[4]

Anthropic estimates that each of these two research studies cost approximately $100,000 in computing power. In the context of global cybersecurity, this is a remarkably small price to pay for identifying fundamental flaws in algorithms that could one day secure trillions of dollars in digital assets.[1][6]

To help the broader research community track AI's evolving capabilities in this highly specialized domain, Anthropic has released "CryptanalysisBench." This new benchmark is designed to evaluate how well various language models can perform cryptographic reasoning and execute formal verification tasks.[1][5]

For enterprise security leaders, the message is clear: there is no need to panic about current encryption standards, but the era of AI-accelerated cryptanalysis has officially arrived. Organizations must maintain visibility over their cryptographic assets to prepare for faster transition cycles.[4][7]

As the world prepares for the massive, multi-year transition to post-quantum cryptography, AI systems will likely serve as an essential proving ground. By breaking candidate algorithms in the lab today, AI is ensuring that the standards securing tomorrow's digital economy are truly unbreakable.[5][6]