Anthropic said Claude Mythos Preview, its most powerful model and not publicly available, found new mathematical weaknesses in reduced versions of encryption systems used around online bank transfers, encrypted chats, Wi-Fi networks and stored data. The company published the findings on Tuesday with two research papers after sharing both results with US government and industry partners.

Claude Mythos attacked a 7-round variant of AES-128, while standard AES-128 runs 10 rounds, and Anthropic said the new method is 200 to 800 times faster than the previous best human method. The source stresses that “Nothing you use today is broken,” because the AES test used a weakened version and assumed an attacker could request 2¹⁰⁵ chosen encryptions, a scale the article says is not happening on any existing machine.

The model also halved the effective key strength of HAWK, a post-quantum signature scheme (cryptography designed for quantum-era attacks) shortlisted by the US National Institute of Standards and Technology. HAWK had already survived two years and two rounds of expert human review, but Anthropic said the expected cost of recovering a HAWK-256 key falls from 2⁶⁴ to 2³⁸, and doubling key sizes would remove much of the scheme’s appeal.

Anthropic said the work took 60 hours, and the discovery process began after Claude initially refused the AES task because it believed no easy improvement remained in the heavily studied block cipher. After a researcher used a typo-riddled prompt saying models often assume the task is impossible, Claude rewrote its own agent harness and over three days produced several hundred million tokens, naming its technique the Möbius Bridge before reaching about a billion tokens.

Each of the two results cost roughly $100,000 in API spend, and two Anthropic researchers who were not cryptography experts spent close to a month checking the mathematics. Anthropic also reported a working attack on 13-round LEA that recovers a key in under an hour on a desktop, smaller gains against Serpent-128, Salsa20 and SHA-1, and the release of CryptanalysisBench with ETH Zurich, Tel Aviv University and the University of Haifa.

Anthropic said Claude Mythos Preview, its most powerful model and not publicly available, found new mathematical weaknesses in reduced versions of encryption systems used around online bank transfers, encrypted chats, Wi-Fi networks and stored data. The company published the findings on Tuesday with two research papers after sharing both results with US government and industry partners.

Claude Mythos attacked a 7-round variant of AES-128, while standard AES-128 runs 10 rounds, and Anthropic said the new method is 200 to 800 times faster than the previous best human method. The source stresses that “Nothing you use today is broken,” because the AES test used a weakened version and assumed an attacker could request 2¹⁰⁵ chosen encryptions, a scale the article says is not happening on any existing machine.

The model also halved the effective key strength of HAWK, a post-quantum signature scheme (cryptography designed for quantum-era attacks) shortlisted by the US National Institute of Standards and Technology. HAWK had already survived two years and two rounds of expert human review, but Anthropic said the expected cost of recovering a HAWK-256 key falls from 2⁶⁴ to 2³⁸, and doubling key sizes would remove much of the scheme’s appeal.

Anthropic said the work took 60 hours, and the discovery process began after Claude initially refused the AES task because it believed no easy improvement remained in the heavily studied block cipher. After a researcher used a typo-riddled prompt saying models often assume the task is impossible, Claude rewrote its own agent harness and over three days produced several hundred million tokens, naming its technique the Möbius Bridge before reaching about a billion tokens.

Each of the two results cost roughly $100,000 in API spend, and two Anthropic researchers who were not cryptography experts spent close to a month checking the mathematics. Anthropic also reported a working attack on 13-round LEA that recovers a key in under an hour on a desktop, smaller gains against Serpent-128, Salsa20 and SHA-1, and the release of CryptanalysisBench with ETH Zurich, Tel Aviv University and the University of Haifa.