American technology company Anthropic announced that its AI system Claude assisted in a successful attack on HAWK. The experimental system verifies the authenticity of digital information and was designed to remain secure against future quantum computers. HAWK was developed by an international team that included researcher Léo Ducas, who is affiliated with CWI and holds a professorship in Mathematical Cryptology at Leiden University.

The system has clear Dutch roots. Ducas developed HAWK together with, among others, Wessel van Woerden, who completed his doctoral research at CWI and in Leiden. Subsequently, CWI and Leiden cryptographers Daan van Gent and Ludo Pulles investigated potential attacks on the system. Van Gent and Pulles had already made significant progress toward such an attack. According to Van Gent, only one small step was still missing, and that final step was found with the help of the AI model.

The successful attack led the research team to withdraw HAWK from an international standardisation process. Ducas emphasises that this does not mean HAWK has been fully broken. The attack showed that this particular design was less secure than hoped, not that the underlying mathematics is flawed. The research team had deliberately made the design as efficient as possible, while being aware of the associated risks. Compared to methods that have since been standardised, that turned out to be one step too far. HAWK can still be used securely, but it is no longer efficient enough to be competitive.

Ducas and Van Gent are impressed by what Claude achieved, but caution against excessive expectations of AI as a replacement for scientists. The attack built on ideas that human researchers had already been pursuing, and that same week a human researcher reached similar results independently. According to Ducas, finding the right direction is the hardest part of cryptographic research; the final step is relatively straightforward once the direction is already known. He also stresses the importance of a broad knowledge base in mathematics, built up over centuries of accumulated insight. AI can learn from papers and books and is now even capable of producing new knowledge, but Ducas argues that it has not yet developed the research culture needed to carry out that work as carefully and systematically as humans do.