Anthropic AI Model Exposes Mathematical Flaws in Next-Gen Quantum and Classical Cryptography

An unreleased artificial intelligence model developed by Anthropic has discovered previously unknown mathematical vulnerabilities in major cryptographic algorithms, including a digital signature candidate currently evaluated for federal standardization in the United States.

The experimental system, designated Claude Mythos Preview, identified an unexploited structural symmetry within HAWK, a post-quantum digital signature scheme designed to secure electronic transactions against future quantum computing threats. HAWK was selected in May as the final lattice-based candidate in the third round of the post-quantum signature evaluation managed by the National Institute of Standards and Technology (NIST).

By capitalizing on the hidden symmetry, the model reduced the computation needed to compromise HAWK’s smallest key configuration from 2^64 operations to 2^38. The shortcut makes key recovery roughly 67 million times less computationally demanding. Addressing the flaw requires engineers to approximately double HAWK’s key size. However, because HAWK was designed primarily around compact key footprints and fast verification—critical attributes for high-throughput networks and blockchain storage overhead—doubling its size removes its competitive advantage over alternative post-quantum schemes.

Anthropic reported that HAWK has not been deployed in active commercial software, and widely used pre-quantum algorithms like ECDSA remain unchanged. The findings were shared with NIST, government representatives, and the algorithm’s developers prior to public disclosure.

Beyond post-quantum systems, the model made notable progress against classical encryption standards. Claude Mythos broke a seven-round research version of the Advanced Encryption Standard (AES-128), accelerating attack speeds by 200 to 800 times compared to the previous academic milestone set by human cryptographers in 2013. Full AES-128 utilizes 10 rounds of cryptographic scrambling to protect global internet traffic, local drive encryption, and financial systems.

To achieve the AES breakthrough, the AI introduced a mathematical technique termed a Möbius Bridge, which effectively removed one of the nine key bytes an attacker was previously required to guess. Researchers initially provided the model with context from roughly 200 unsuccessful prior attempts and restricted it from using five traditional cryptanalysis approaches. Although the AI initially concluded that no vulnerabilities were reachable, researchers directed the system over three days to continue exploring novel attack vectors.

In a third demonstration, the model compromised a 13-round variant of LEA—a lightweight encryption scheme recognized as an ISO standard and South Korean national benchmark for Internet of Things devices—in under an hour using a standard desktop computer. Previous academic efforts against 13-round LEA required 2^98 plaintext pairs. Fully deployed LEA uses 24 rounds and remains secure.

The mathematical discoveries were largely generated by the AI, with human researchers primarily directing targets, structuring experiments, and verifying mathematical proofs. Anthropic staff spent several hundred hours confirming the validity of the AES attack. Generating each cryptographic result required approximately $100,000 in API compute resources.

The labor-intensive verification phase underscores an emerging challenge within cybersecurity. Having previously identified 271 software bugs in Mozilla Firefox during internal trials, Anthropic noted that automated AI models are producing novel security findings at a rate that threatens to overwhelm traditional human verification, triage, and patching pipelines.

To evaluate these capabilities systematically, Anthropic established CryptanalysisBench, a benchmark suite containing 191 automated cipher-breaking challenges derived largely from NIST competitions. An updated iteration of the model, Mythos 5, successfully solved 85.7 percent of tasks with known mathematical solutions, compared to 65.3 percent for the lowest-performing model tested. When assigned full-strength, operational ciphers with no existing public vulnerabilities, all models evaluated scored under 9 percent success.