Anthropic's Claude Mythos Preview has independently discovered improved attacks on two major cryptographic algorithms, signaling that frontier artificial intelligence models can now perform expert-level cryptanalysis work that previously required specialized human researchers. The model found mathematical weaknesses in HAWK, a post-quantum signature candidate under review by the National Institute of Standards and Technology (NIST), and in a reduced-round variant of AES-128, the world's most widely used symmetric encryption standard. Each discovery cost approximately $100,000 in API usage, yet the findings took only days for the AI to produce compared to years of human research.

What Makes These Cryptography Discoveries Significant?

The timing of these findings matters considerably. Earlier this year, Google DeepMind used Gemini to resolve open mathematical problems, and OpenAI's GPT disproved the unit distance conjecture. Now Anthropic's own Fable 5 resolved the Jacobian Conjecture. Cryptography represents the next frontier, but unlike pure mathematics, the stakes are immediately practical because the algorithms under scrutiny protect real data in production systems worldwide.

HAWK is one of the remaining candidates in NIST's search for post-quantum digital signatures, algorithms designed to remain secure even against cryptographically relevant quantum computers. It survived two rounds of expert review over two years. Mythos found an improved attack in roughly 60 hours by exploiting a previously undetected symmetry in the mathematical lattice that HAWK's security relies on. The attack effectively halved the scheme's key strength: the expected cost of a full key recovery against the small HAWK-256 parameter dropped from 2 to the 64th power to 2 to the 38th power.

The second result targets AES-128, the cipher adopted by NIST in 2001 that has been scrutinized more intensively than almost any encryption algorithm in existence. Mythos attacked a reduced-round variant with 7 rounds instead of the full 10, which is a standard academic technique for probing a cipher's security margins. The model developed a novel fingerprinting method Anthropic named "Möbius Bridge," which works by making a certain mathematical transformation invariant to one of the guesses an attacker normally has to enumerate. Combined with several additional optimizations Mythos discovered, the attack improves on the best previously known meet-in-the-middle attacks on 7-round AES by a factor of 200 to 800.

How Did the AI Actually Discover These Flaws?

• Multi-Agent Collaboration: The HAWK attack involved a multi-agent workflow where one worker agent initially dismissed the idea as infeasible, but a second agent found a way to fully exploit the symmetry. The pair kept exchanging messages until both agreed they had a viable attack.
• Persistent Problem-Solving: Mythos initially refused to work on the AES problem, telling the researcher that "if you want a different outcome, the target has to change." It only began pursuing creative approaches after the researcher pushed back with messages emphasizing the need for genuine research breakthroughs.
• Computational Scale: The AES attack generated approximately one billion output tokens over three days, with the human researcher primarily handling project management rather than performing the cryptanalysis directly.

The approximately one billion output tokens Mythos generated for the AES attack cost roughly $100,000, or about $0.10 per thousand tokens for Mythos output. While that may sound expensive, compare it to the alternative: two human researchers spending several hundred hours just to validate what the model produced. The researchers who verified the AES result were not cryptography experts and had to learn enough of the field to check the work, a process that took them nearly a month.

What Other Cryptographic Systems Did Mythos Attack?

Anthropic also disclosed additional findings that are still being analyzed. These include a practical key-recovery attack on 13-round LEA, a lightweight cipher standardized in ISO/IEC 29192-2:2019, that runs in under an hour on a desktop computer. The company also reported attacks on round-reduced variants of Serpent-128, Salsa20, Poseidon, and SHA-1. Most are marginal improvements, but the LEA attack is notable because it is end-to-end verifiable: you pick a random key, run the attack, and recover it.

A separate benchmark called CryptanalysisBench, co-developed by ETH Zurich alongside researchers from Tel Aviv University and the University of Haifa, packages 191 cryptanalysis tasks across six primitive families: block ciphers, hash functions, stream ciphers, AEAD schemes, message authentication codes, and public-key encryption. Five frontier models, including Claude Opus 4.8, Sonnet 5, Mythos 5, GPT 5.5, and the open-weights GLM 5.2, already break 65 to 86 percent of tier-1 schemes, and models are producing novel cryptanalysis, including a previously unknown key-recovery attack against the SpoC AEAD scheme.

What Are the Real-World Implications?

Anthropic was careful to state that neither finding affects production systems. HAWK is only a NIST candidate, not deployed anywhere. The AES attack applies to a reduced-round variant, and breaking the full 10-round cipher remains out of reach. Even if it were possible, the attack would cost hundreds of millions of dollars to execute under its theoretical model.

However, the research raises important questions for European cybersecurity and regulatory frameworks. Anthropic Ireland Limited is the legal entity providing Claude services in the European Union, operating under EU jurisdiction and supporting all member states. Services marketed in the EU must comply with the General Data Protection Regulation (GDPR) and, increasingly, the EU AI Act. Cryptographic security underpins GDPR's data protection requirements. If the algorithms that protect personal data turn out to have mathematical weaknesses, that becomes a compliance issue.

The EU AI Act classifies AI systems used in critical infrastructure as high-risk. Cryptographic libraries certainly qualify. The question Anthropic's research raises is whether an AI model that can find vulnerabilities in those libraries, or in the algorithms themselves, should be treated differently under the Act, especially if it is restricted from public access. Mythos Preview remains unavailable to the public.

This research is an early warning for European cybersecurity agencies. The EU Agency for Cybersecurity (ENISA) already coordinates vulnerability disclosure across member states. If frontier AI models can autonomously discover flaws in the mathematical design of cryptographic primitives, the volume of findings could overwhelm the existing disclosure pipeline. Anthropic itself notes that the bottleneck has shifted: it took one week for Mythos to discover the AES attack, but a month for humans to confirm it.

" }