A post-quantum cryptography (PQC) algorithm under consideration for a future US cybersecurity standard has been withdrawn after Anthropic's AI security model, Mythos, helped uncover a critical flaw that significantly weakened its security.
The algorithm, known as HAWK, was one of the digital signature schemes being evaluated by the US National Institute of Standards and Technology (NIST) as part of its effort to standardise encryption capable of resisting future quantum computer attacks.
Following Anthropic's disclosure of the findings, HAWK's developer announced the algorithm would be withdrawn from the NIST evaluation process.
AI-assisted cryptanalysis
Anthropic said its Mythos model worked semi-autonomously for around 60 hours, using approximately $100,000 in computing resources to improve an existing attack against HAWK.
Rather than inventing entirely new mathematics, Mythos combined known cryptographic techniques in a previously unexplored way, effectively reducing the algorithm's security strength by about half.
In cryptography, an algorithm is generally considered "broken" when attackers can recover secret keys more efficiently than by brute force.
The flaw could technically be mitigated by doubling HAWK's key size, but doing so would make the algorithm less efficient than competing post-quantum signature schemes already available.
Cryptography expert Matthew Green of Johns Hopkins University described the discovery as significant because it creatively combined existing methods that researchers had not previously connected.
Google post-quantum cryptography expert Sophie Schmieg said the result effectively ended HAWK's chances of becoming a standard.
"Basically, HAWK is dead," she said.
AES research less dramatic
Anthropic also reported improvements to a known attack against the widely used Advanced Encryption Standard (AES).
Using Mythos, researchers refined a "meet-in-the-middle" attack on a deliberately weakened version of AES, reducing the computational effort required under laboratory conditions.
However, experts stressed that the findings do not threaten the security of AES used in real-world systems. The experiments were conducted on simplified challenge versions designed specifically for research rather than production-grade encryption.
No immediate risk to users
Security researchers emphasised that neither discovery places current internet encryption at immediate risk.
The attacks were demonstrated only against weakened testing versions of cryptographic systems, and both remain impractical outside controlled research environments.
Nevertheless, the findings suggest artificial intelligence could become an increasingly valuable tool for identifying weaknesses in future cryptographic designs before they are adopted as security standards.
Anthropic said AI-assisted cryptanalysis could eventually accelerate vulnerability discovery faster than human researchers can independently validate and address new findings.
Growing role for AI in cybersecurity
The withdrawal of HAWK highlights the growing role of AI in cybersecurity research, where advanced language models are beginning to assist experts in analysing complex mathematical problems.
While researchers caution against overstating AI's capabilities, they say the latest results demonstrate that AI could increasingly contribute to testing and strengthening next-generation encryption technologies as governments and industry prepare for the era of quantum computing.
