MiniMax says an internal-only AI model has developed working attacks on academic cryptographic designs and past competition candidates. Its examples range from recovering shared secrets in seconds to exploiting a signature-verification bug caused by a missing ampersand. Mathematics, meet proofreading. Source
MiniMax says this work does not affect deployed cryptographic standards. Its targets are academic designs, past competition candidates and public challenge parameters. Source
Attacks on the math and the code
The company calls its research model MiniMax’s Internal Model, or MIM. For LOCKER, a past NIST post-quantum cryptography candidate, MiniMax says MIM attacked the underlying mathematical problem after buffer bugs in the reference code were patched. It reports reducing the solver matrix from 42 petabytes to 58 megabytes and recovering the shared secret across all parameter sets in two seconds. Source
The two-second figure is the reported recovery time. The matrix reduction isn’t a miraculous compression format. Source
In the DRS signature scheme’s reference code, MiniMax says MIM found logical AND (&&) mistyped as bitwise AND (&). It says the bug allowed forged signatures using only the public key. That is an implementation flaw, distinct from breaking the mathematical design. Source
By contrast, MiniMax reports a structural flaw in COLM, a winner of the CAESAR authenticated-encryption competition. It says MIM produced a forged encrypted message that passed verification using just two requests to encrypt chosen messages, with no requests to decrypt ciphertexts. Source
Big speedups, specific targets
MiniMax also says MIM recovered secret keys for 94 TII McEliece challenge instances, including 55 previously unsolved hard instances. On instances with prior public records, it reports runtimes of 1.80–5.36 seconds on a single 64-core server, describing speedups of 6,180 to 43,400 times against earlier records and Anthropic’s multi-GPU solver. Source
Those speedups compare selected challenge instances across different hardware, not the models’ overall performance. They also do not demonstrate a break of production-strength Classic McEliece. Source
Working attacks still need expert scrutiny
MiniMax says it shared the findings with the respective scheme authors in advance and released technical reports and proof-of-concept code through its Research GitHub repository. Source
MiniMax writes: “Working exploit code shows that an attack succeeds, but evaluating the security margins of cryptographic designs still depends on review by human cryptographers.” Source
MIM itself remains internal-only, although MiniMax says qualified researchers and developers can apply for testing access. Source
1 comment