DeepMind finds two AI-assisted weaknesses in 1024-bit RSA

Google DeepMind says AI models, working under the guidance of outside cryptography researchers including Nadia Heninger and Emmanuel Thomé, found two new ways to attack legacy 1024-bit RSA keys faster than any previously published method. The work appeared in two papers published on 5 and 6 November, one on forging digital signatures, the other on factoring many keys sharing a batch.

The signature-forgery method is roughly 1,600 times more efficient than the previous best technique, in cases where an attacker controls part of the signed data, such as a short digital fingerprint. The batch-factoring method cuts the marginal cost of cracking each additional 1024-bit key in a large batch by about 30 times. DeepMind says both results came out of an isolated research environment and that RSA keys of 2048 bits or longer are not affected.

Why it matters: 1024-bit RSA has been deprecated for new certificates for over a decade, but it still turns up in old DKIM signing keys, legacy SSH host keys and forgotten internal certificates. If you’ve never audited for it, this is a reason to: find any 1024-bit RSA keys still in use and rotate them, rather than assuming they’re too old to be worth attacking.

The caveat: these are theoretical estimates from a controlled research setting, not a demonstrated attack against a live system, and DeepMind hasn’t published full reproduction details alongside the headline figures.