Anthropic says Mythos found new weaknesses in test encryption

In short: Anthropic says its Mythos AI security model found new ways to attack two weakened, test versions of encryption systems, including one meant for a future with quantum computers.

What happened

Anthropic reported results from two cryptography tests it ran with its Mythos model, which is only available to a small group of trusted users. Cryptography is the math that helps keep messages private and proves who signed a file, like a tamper proof seal on a document.

In the first test, Mythos helped improve an existing attack on HAWK, a digital signature design being evaluated by NIST, a US standards body. A digital signature is a way to prove a message really came from a person or company and was not changed (like a wax stamp, but for data). Anthropic said the improved method effectively cut HAWK’s key strength in half, after about 60 hours of work and around $100,000 in computing costs, even though the human operator was not a cryptography expert.

Importantly, this was not the full, real world version of HAWK. It was a weaker “challenge instance” made for testing, and the attack would likely be impractical outside a lab. Even so, experts quoted by Ars Technica said the result could make HAWK less competitive, since it may need larger keys to stay safe.

In the second test, Mythos improved a known research attack on a weakened form of AES, a widely used encryption standard. Anthropic said the improvement reduced the amount of chosen test data needed, but the test used a reduced round version of AES, not the standard one used in products.

Why it matters

These results do not mean today’s common encryption is broken. They do suggest AI tools may help researchers find weak spots faster, which could change how new security standards are tested before they are widely adopted.

Source: Arstechnica