Realtime AI News
MiniMax Internal Model Breaks Multiple NIST Candidate Ciphers, Cutting Cryptanalysis to Seconds
On November 22, MiniMax published cryptanalysis results produced by MIM, an internal research model that ran systematic automated attacks on modern cryptography, surfacing previously unknown construction flaws and turning paper-only breaks into executable attacks that finish in seconds to minutes. MiniMax said the work targets academic designs, historical competition candidates and public challenge parameters, not deployed standards such as AES, RSA or SHA-256.

On the evening of November 22, Shanghai-based model developer MiniMax published cryptanalysis results produced by MIM, an internal research model it says ran systematic, automated analysis against modern cryptographic schemes. According to the company, the model surfaced several previously unknown design-level flaws and pushed a number of attacks that had only existed on paper into executable breaks that finish in seconds to minutes.
MiniMax grouped the findings into three tracks. Against candidates from past NIST cryptographic standardization competitions, MIM delivered the first practical breaks of several algorithms for which no known practical attack had previously existed. Against schemes published at top cryptography conferences and journals, or winning competition designs, it produced efficient authenticated-ciphertext forgery and key recovery, and set a new record for the number of rounds broken in a recently designed block cipher. It also attacked the classical hard mathematical problems that public-key cryptography rests on, combining exploitation of hidden algebraic structure with low-level engineering work.
Concrete cases are more telling than the summary. On Waterfall-256, a SHA-3 competition candidate, MIM built the first practical collision attack under blind testing, finding collisions in 0.045 seconds on a single core; on Blender-256 it achieved a full-round collision break that had not been accomplished since the algorithm was submitted. Blind testing matters here: the model was not handed the target structure in advance, so the attack had to be searched out.
Post-quantum schemes were a major focus. MIM broke the rank-metric decoding problems behind LOCKER and ROLLO-I, compressing the matrix required for the former from 42 petabytes to 58 megabytes and recovering the shared key within two seconds, while recovering an equivalent private key for the latter from the public key in roughly six seconds. The model also found more than a dozen unreported implementation flaws in reference code for DRS, RQC-128 and LAC-128.
For designs that had already passed peer review, MIM's wins landed at the level of security proofs. It built a forgery attack on COLM, the CAESAR authenticated-encryption winner, without any decryption queries; for the post-quantum signature scheme SBC-VOLE, a single known signature was enough to extract the full private key in 0.06 seconds; and it extended practical key recovery on ARADI, an NSA low-latency cipher, from five rounds to eight.
The hardest problems produced the largest numbers. In the TII McEliece public challenge, MIM recovered keys for 94 instances, 55 of them previously unsolved, at speeds 6,180 to 43,400 times faster than the best existing solvers. Its estimates also cut the claimed security strength of the HAWK signature scheme by 9.3 to 23.8 bits.
MiniMax stressed that the work targeted academic proposals, historical competition candidates and public challenge parameters, and does not affect deployed standards such as AES, RSA or SHA-256. Following responsible disclosure practice, the company said it notified the authors of the affected schemes in advance, and it is releasing the model's reasoning traces as a technical report together with proof-of-concept code on GitHub.
The result matters less as a set of records than as evidence that frontier models can now run end-to-end research loops in a highly technical field: proposing attacks, testing them against real implementations, and iterating until they work. It also puts a question to the cryptography community — if automated tools can sift through decades of candidate designs in seconds, peer review and public testing practice may need to move at a similar pace.
Why it matters
MiniMax has handed a narrowly technical research task to an internal model and attached open reports and code to it, giving AI4S a rare verifiable sample. For the security field, maturing automated cryptanalysis means many long-unexamined candidate designs are likely to be re-evaluated quickly.
Nearby Updates
All11/23, 01:21
Genspark releases VoicePad, a five-key AI voice control device
Genspark has released VoicePad, an AI voice-control device built around just five physical keys, according to Chinese tech news channels. Details remain thin beyond the product name and form factor, with no pricing, ship date, or feature list disclosed.
11/23, 01:19
Report: Anthropic asks IPO underwriters to accept six-month non-compete
Anthropic is reportedly requiring underwriters on its IPO to accept a six-month non-compete, according to Chinese tech news channels. The disclosure covers only that term, with no syndicate list, valuation range, or offering timetable made public.
11/23, 00:30
Australia Puts AI Safety at the Heart of Foreign Policy in Wong Speech
Australia will place AI safety at the centre of its foreign policy, with Foreign Minister Penny Wong set to tell parliament on Tuesday that guarding artificial intelligence must become a pillar of the country's diplomacy. The speech previews national AI standards due by year's end and legislation in 2027, alongside unresolved fights over copyright and tech regulation.
11/22, 22:34
Tencent Publishes Safety Report for Its Hy4 AI Model
Tencent has published a safety report for Hy4, its newest AI model, setting out the company's own assessment of how the model behaves on safety-related tests. The disclosure puts Tencent's evaluation approach on the record, though the specific scores and testing methodology have yet to be detailed publicly.