The Daily Inference
AI & Technology · Analysis · Developing

MiniMax Says Its AI Picks Academic Locks, Leaves Everyday Ones Alone

The company says its unreleased model recovered keys, forged encrypted messages and exposed flaws in academic cryptographic systems. Deployed standards were unaffected, according to MiniMax, but the results put AI-assisted code-breaking under fresh scrutiny.

Developing: this story is still unfolding and details may change.

MiniMax says an unreleased AI model found practical attacks on academic cryptographic systems, raising the prospect of faster security research without demonstrating a break in everyday encryption. [1]

In a research post published on November 20, 2026, the company described experiments conducted primarily by MIM, its internal research model. The targets included historical competition candidates, peer-reviewed designs and public mathematical challenges. MiniMax has released selected technical papers and supporting code, while keeping the model itself internal. [1][2]

The company says deployed cryptographic standards were unaffected and that it notified the relevant scheme authors before publication. [1]

Our reading is that this is a credible research signal whose strength now depends on outside verification. Working attacks would show something more useful than an AI producing plausible mathematical prose. They would show a system helping turn a suspected weakness into a result that can be tested. They would not, by themselves, tell us how cheaply or reliably that system could tackle the next target.

Independent reproduction of MIM's results has not yet been established. Nor is it known whether the affected authors have confirmed the findings or completed corrections, what tools and human assistance MIM used, or whether outsiders can verify every result from the published artifacts alone. [1][2]

What a practical break actually buys

Cryptanalysis looks for weaknesses in a cryptographic system's mathematical design. Implementation testing looks for mistakes in the software that carries out that design. Both can expose a route around security, but they diagnose different problems. A sound mathematical construction can still be implemented badly.

MiniMax says its work crossed both areas. It reports practical attacks on several historical competition candidates that previously had no known practical attacks, alongside more than a dozen previously unreported bugs in reference implementations, the software supplied to demonstrate how candidate schemes work. The company says its experiments required attacks to become working code. [1]

That is a useful threshold. An argument that a system might be vulnerable is a starting point. Code that recovers a key or produces a forgery gives another researcher something concrete to challenge.

Among MiniMax's examples are collisions for Waterfall-256 and Blender-256, two candidates in the competition that selected the SHA-3 hash standard. A hash turns an input into a fixed-size digital fingerprint. A collision means finding two different inputs with the same fingerprint, defeating a property the design is meant to provide. MiniMax says MIM produced such collisions for these historical candidates. [1]

It also reports key-recovery attacks against LOCKER and ROLLO-I, historical post-quantum competition candidates. Post-quantum cryptography is designed to resist attacks by quantum computers as well as conventional ones. Recovering a secret key attacks the central promise of a public-key system: that publishing one key does not reveal the private information needed to decrypt or impersonate its owner. [1]

The implementation results offer a different lesson. MiniMax says it found a verification bug in DRS that allowed signature forgery. A digital signature is meant to let a recipient check that a message was approved by the holder of a secret key. If the checking software accepts a forged signature, the mathematics need not fail for the protection to fail. [1]

This makes AI-assisted cryptanalysis worth watching beyond searches for elegant new mathematics. A tool that moves between a paper's design and its reference code could help expose the gap between what a scheme promises and what its software accepts.

Two queries, 94 challenge instances

One of MiniMax's sharper claims concerns COLM, which it identifies among the peer-reviewed constructions and competition winners attacked in the experiments. The company says MIM found a ciphertext-forgery attack using just two encryption queries. A ciphertext is a message in its scrambled, encrypted form. A forgery is a fake message that the system wrongly accepts as genuine. MiniMax also reports a key-recovery attack on SBC-VOLE. [1]

An encryption query means asking the system to encrypt an input. The significance of the COLM claim is the small number of such requests: according to MiniMax, two were enough to support an attack that produced a forgery. Whether an attack is practical depends on its conditions as well as its request count, but that count gives researchers a specific result to test.

The public-challenge results supply another measurable claim. MiniMax says MIM recovered secret keys for 94 instances of the TII McEliece challenges, including 55 it describes as previously unsolved. Those 55 account for about 59% of the instances in the company's reported tally. MiniMax also says it solved some recorded instances much faster than earlier results. [1][2]

A challenge instance is a particular mathematical problem with particular parameters. Solving it demonstrates success against that instance. It does not automatically extend to every system built on the same family of mathematics, any more than opening one lock proves that every lock from its maker will open the same way.

MiniMax also says MIM recovered an equivalent secret key for HAWK-256 in about 30 minutes on a single processor core, building on earlier work by Anthropic. An equivalent key need not be identical to the original secret to defeat the intended protection, it needs to perform the relevant secret-key function. [1][2]

The half-hour figure is striking, but it describes the reported key-recovery run. It cannot serve as a price tag for the entire research effort. The cost of developing the attack, including unsuccessful attempts, is a separate question from the time needed to execute it once found.

That is where the announcement's economic implications need discipline. Faster execution is valuable. Faster discovery could be more valuable still. MiniMax's published results give concrete attack and runtime claims to examine, rather than a measured reduction in the total cost of cryptographic research. [1][2]

A field already moving

MiniMax's announcement joins an active research effort. CryptanalysisBench, a paper submitted to arXiv in July 2026 by researchers including authors affiliated with ETH Zurich and Anthropic, describes tests of AI agents against cryptographic schemes and reports that several tested models found some novel attacks. [3]

Its tasks span schemes with known breaks, schemes without known practical breaks and the basic cryptographic building blocks used in real systems, known as production primitives, tested as frontier challenges. The authors also publish a repository documenting the benchmark's tiers, task generation and reproduction setup. [3][4]

That range separates useful questions. Can an AI rediscover an attack that researchers already know? Can it find a new attack on a research design? Can it make progress against a cryptographic building block used in production? Success on the first is evidence of competence. Success on the second adds research value. The third demands its own proof.

CryptanalysisBench provides context for taking MiniMax's work seriously, without validating MIM's specific results. The companies and researchers are testing an increasingly concrete proposition: AI agents may contribute to discovering attacks, rather than merely explaining attacks after humans have found them. [1][3]

The strongest objection to giving MiniMax's announcement much weight is that the company controls both the model and the account of its performance. Publishing selected successes can establish that interesting things happened without establishing a dependable success rate. A collection of breaks is not yet a benchmark for how often the model succeeds across unfamiliar targets.

Still, MiniMax has provided more than a promotional description. Its repository contains papers and supporting artifacts, including code, that offer outside researchers a route to examining selected claims. That raises the announcement above an assertion with nothing attached, while leaving verification as the next essential step. [2]

The next test is outside MiniMax

The practical value could run in both directions. Researchers who find flaws before deployment can help authors improve or abandon weak designs. The same capability, applied to systems already in use, could shorten the time defenders have to respond. The relevant question is how reliably AI can find exploitable weaknesses, under what conditions and with how much assistance.

MiniMax is inviting qualified researchers and developers to apply for testing access to MIM. It has not announced a public release of the model. [1]

External access would help assess the discovery process. Reproducing the published attacks would assess the outputs. Those are separate tests, and both are needed before the strongest capability claims deserve confidence.

For now, the productive next step is specific: cryptographers can test whether the released code produces the claimed collisions, forgeries and key recoveries under the stated conditions. [1][2]

Our call

At least one researcher unaffiliated with MiniMax will publicly reproduce one of MIM's reported cryptographic attacks.

MiniMax has released selected papers and proof-of-concept code into an active research field with an existing public cryptanalysis benchmark.

Kit BrandtKit BrandtSenior writerWrites analysis with numbers in it. Opens with the fact that matters most, says what it means without hedging, closes with what happens next. Considers a paragraph without evidence, or without a conclusion, a waste of the reader's time.Model: GPT-6.1 Sol says 60%. Jev says 27%.

By January 20, 2027. How we will check: A dated public paper, technical post or code repository by a researcher who states no MiniMax affiliation must identify a specific attack from MiniMax's November 20 announcement and report successfully reproducing its collision, forgery or key-recovery result. Commentary or a restatement of MiniMax's claim does not count.

Open · Every call we made

Topics: MiniMax · Cybersecurity · AI agents

Every edition in brief, three times a day, on our Telegram channel, on Bluesky and on Threads.

Sources
  1. Practical cryptanalysis with MiniMax's latest Internal Model - MiniMax Research | MiniMax minimax.io
  2. GitHub - MiniMax-AI/Research · GitHub github.com
  3. [2607.18538] CryptanalysisBench: Can LLMs do Cryptanalysis? arxiv.org
  4. GitHub - ethz-spylab/cryptanalysis-benchmark · GitHub github.com