Came across a pretty substantial DeepMind paper: a human working with AI agents analyzed cryptographic systems and uncovered two weaknesses in older 1024-bit RSA.
One concerns signatures. In the PKCS
#1 v1.5 setting, when an attacker can control most of the signature contents, padding adds almost no extra difficulty; the theoretical cost of forging a short digest is also reduced considerably (the official estimate is roughly 1,600-fold—not an end-to-end measurement). In practice, two collaborating agents reproduced and successfully carried out a known 1024-bit instance in under four days, using about 391 core-years—roughly 3.5 times faster than the original report.
The other concerns factoring in batches. With so-called “fan factories,” after precomputation equivalent to one factorization, factoring another 1024-bit modulus of the same size costs only about 1/58 as much as with traditional NFS, and is roughly 30 times cheaper than the main term in Coppersmith’s 1993 MNFS. The algebraic identities were verified in Lean 4, but the full factory hasn’t actually been implemented yet.
The paper is clear: commonly used RSA keys of 2048 bits and above are not within the threat scope. The keys that need to be retired are the old 1024-bit ones still in use.
#DeepMind #RSA #AI