Claude isn't just coding, it's driving scientific discovery

Can an AI be more than just a code generator, evolving into a true collaborator in highly specialized scientific and engineering domains? Recent breakthroughs involving Anthropic's Claude suggest this is not just possible, but already happening, marking a significant shift in how we conceive of AI's role in cutting-edge technical work.

Anthropic’s research model, Claude, recently achieved a remarkable feat: tracing the origins of the NIST ECDSA curve seeds to a 1990s C library's lrand48 random-number generator, seeded with Unix timestamps. This answered a provenance question that had baffled cryptographers and security researchers since the NSA provided the values without explanation in 2000. What's more, Claude achieved this in under half an hour, disproving a long-standing theory about a humorous passphrase and demonstrating a methodical, generator-side approach to problem-solving. This isn't just about finding an answer; it's about solving a complex cryptographic riddle that human experts had struggled with for decades. Claude’s method involved recognizing that the lrand48 generator returns only the top 31 bits of a 48-bit state. It used Lenstra–Lenstra–Lovász (LLL) basis reduction to reconstruct the full internal state, identifying Unix timestamps between May 1997 and April 1998 as the initial seed values for all ten prime-curve seeds. Additionally, Claude traced the base points for several curves to the random number generator in FreeLIP, Arjen Lenstra’s long-integer arithmetic package. Anthropic emphasizes that these findings do not represent a cryptographic break or a weakness in the ECDSA curves, but rather an elucidation of their origins.

Beyond historical provenance, Claude is also making strides in forward-looking fields, demonstrating its capability to discover actual mathematical flaws in cryptographic algorithms. Using Claude Mythos Preview, researchers at Anthropic discovered improved ways to attack a candidate post-quantum digital signature scheme called HAWK, effectively halving its key strength in just 60 hours. HAWK had previously survived two years of expert human review. In another instance, Mythos found a new way to attack a round-reduced version of AES, the Advanced Encryption Standard, improving the speed of previous best attacks by 200-800x. While neither of these findings currently affects production systems — HAWK is still a candidate and the AES attack is on a weaker version — they highlight the potential for frontier AI models to uncover vulnerabilities in crucial cryptographic algorithms, both before and after real-world deployment. These results underscore AI's role in cryptographic research, stress-testing algorithms to build more robust security systems.

We've also seen Claude contribute to optimizing quantum circuits.