AI Finds New Weaknesses in Cryptographic Algorithms, Anthropic Says

Insider Brief

• Anthropic says its Claude Mythos Preview AI model discovered improved mathematical attacks against the HAWK post-quantum digital signature candidate and a reduced-round version of AES, demonstrating that advanced AI can contribute to cryptography research without affecting current production systems.
• The AI reduced HAWK’s effective security margin by identifying a previously unknown mathematical symmetry and developed a new attack that is 200 to 800 times faster against a research version of seven-round AES, while researchers spent weeks validating the model’s findings.
• Anthropic says the results highlight AI’s growing role in evaluating encryption algorithms before deployment, prompting the company to release new research papers and CryptanalysisBench to help researchers measure AI capabilities in cryptanalysis.

Anthropic says one of its newest AI models discovered improved mathematical attacks against two important cryptographic algorithms, demonstrating that frontier AI systems may soon become valuable tools for testing the security of the encryption methods that protect digital communications.

The company said the findings do not threaten today’s deployed encryption systems. Instead, they strengthen the process of evaluating cryptographic designs before they are widely adopted, according to an Anthropic research blog post.

The research marks a shift from AI identifying programming mistakes in software implementations to finding weaknesses in the underlying mathematics of cryptographic algorithms themselves.

According to Anthropic, its Claude Mythos Preview model developed an improved attack against HAWK, a candidate digital signature algorithm being evaluated as part of the National Institute of Standards and Technology’s post-quantum cryptography effort. The model also devised a faster attack on a reduced-round version of the Advanced Encryption Standard, or AES, the encryption algorithm used worldwide to protect internet traffic, financial transactions and other sensitive communications.

Anthropic emphasized that neither result requires changes to production software.

HAWK has not been standardized or deployed, while the AES work applies only to a deliberately weakened research version of the cipher rather than the full algorithm used in practice.

AI Moves Beyond Software Bugs

The findings build on Anthropic‘s earlier demonstrations that Claude could identify security flaws in software code. Those earlier vulnerabilities stemmed from implementation mistakes made by programmers.

This latest work is different because it targets the mathematical foundations of cryptographic systems rather than coding errors, according to the company’s researchers.

Cryptographic algorithms underpin much of modern digital security. Digital signature schemes verify that users are communicating with authentic websites, while symmetric encryption algorithms such as AES protect data as it travels across networks. Weaknesses in these systems could expose communications, banking transactions and other online activity to attackers.

Anthropic said the research illustrates how advanced AI models may increasingly assist cryptographers by identifying weaknesses during the design and evaluation of encryption systems rather than after they have been deployed.

The company also said it followed responsible disclosure procedures throughout the project, consulted outside academics to validate the results, shared advance copies with government and industry partners and coordinated disclosure with the HAWK developers before releasing the findings publicly.

HAWK Security Margin Reduced

HAWK is one of the remaining candidates in NIST’s competition to identify digital signature schemes that can resist attacks from future quantum computers.

Quantum computers are expected to eventually threaten widely used public-key cryptography systems such as RSA and ECDSA by solving mathematical problems that are impractical for conventional computers. NIST has spent nearly a decade evaluating replacement algorithms that can withstand those future attacks.

According to Anthropic, Claude Mythos Preview discovered a previously unexploited mathematical symmetry within HAWK’s underlying lattice structure. That discovery enabled a faster method for recovering encryption keys than researchers had previously identified.

The result effectively cuts HAWK’s security margin in half by reducing its effective key size, according to the company. Anthropic said larger HAWK key sizes would remain impractical to attack, but maintaining equivalent security would require doubling key sizes, eliminating many of the performance advantages that made HAWK an attractive post-quantum candidate.

The company reported that discovering weaknesses during the NIST review process is expected. Several previous candidates in the agency’s post-quantum competition were ultimately found to be insecure before standardization, including one algorithm that researchers completely broke in roughly an hour using a laptop.

Anthropic said Claude required about 60 hours to develop and verify the HAWK attack while working in a multi-agent environment with limited human guidance. The company estimated the effort consumed roughly $100,000 in API usage. Human involvement primarily focused on project management and computational verification rather than cryptographic expertise, according to the blog.

Faster Analysis of Reduced AES

The second result focused on AES, the symmetric encryption standard adopted by NIST in 2001 and one of the most extensively analyzed cryptographic algorithms in existence.

Rather than attacking the full 10-round AES-128 algorithm used in production, Claude targeted a seven-round research version. Cryptographers routinely study reduced-round versions because they are easier to analyze and can provide insight into the security margins of the complete cipher.

Anthropic said Claude improved upon the best previously known “meet-in-the-middle” attack, a cryptanalytic technique that speeds computation by storing intermediate calculations and reusing them later. The model developed what it called a “Möbius Bridge,” a new fingerprinting method that eliminated one stage requiring 256 separate guesses.

According to the company, the improvement accelerated previous attacks by between 200 and 800 times, depending on how runtime is measured. Even so, the attack remains impractical because it assumes an attacker can obtain approximately 2¹⁰⁵ chosen plaintext encryptions, a scenario used for academic analysis rather than real-world attacks.

Anthropic said the model initially resisted the project, repeatedly concluding that meaningful improvements were unlikely because AES has already received decades of scrutiny.

Researchers then modified Claude’s prompting to encourage it to search for genuinely novel ideas rather than incremental improvements.

Over the following several days, the model generated hundreds of millions of tokens while autonomously developing hypotheses, testing ideas and refining the attack. Anthropic researchers later spent several hundred hours validating the mathematical correctness of the result and preparing the accompanying research paper.

Researchers Face a New Bottleneck

Anthropic said the work suggests AI may soon shift the limiting factor in cryptography research from discovering new attacks to verifying them.

The company said Claude developed the improved AES attack in about one week, while two researchers required nearly a month to gain confidence that the method was mathematically correct.

Anthropic said it has already expanded its research beyond HAWK and AES. According to the company, Claude has developed a practical attack against a reduced-round version of the Lightweight Encryption Algorithm, or LEA, and another full key-recovery attack against a reduced-round version of the Serpent-128 cipher. It has also identified smaller improvements involving Salsa20, the Poseidon hash function and SHA-1, although those results remain under investigation. None of those findings affect the full production versions of the algorithms.

To help other researchers measure AI performance in cryptanalysis, Anthropic partnered with researchers at ETH Zurich, Tel Aviv University and the University of Haifa to create CryptanalysisBench, a benchmark designed to evaluate language models against a range of cryptographic challenges.

In the future, Anthropic said it expects AI systems to play an increasingly important role in reviewing new cryptographic designs before they are deployed, while also raising questions about how governments, industry and academia should respond if AI eventually discovers vulnerabilities that affect systems already protecting real-world data.

The company said adversarial review has always been central to cryptography, with researchers continually testing proposed algorithms until the community is satisfied with their security. Anthropic argued that highly capable language models are likely to become another participant in that process, helping strengthen future encryption standards while also increasing the volume of research that human experts must evaluate.

Read the full paper on HAWK.

Read the full paper on AES, and the associated chain of thought.

Read the paper introducing CryptanalysisBench.