Researchers have introduced CryptanalysisBench, a benchmark of 191 tasks across six families of cryptographic primitives, to test how well large language models can perform cryptanalysis. The results point to a useful new security question: five frontier models broke between 65% and 86% of the benchmark’s Tier 1 schemes, according to the project’s publication page.
Watch Desk analysis
What happened
The benchmark is designed to measure models’ ability to analyse cryptographic schemes, and to help stress-test candidate schemes before deployment. The researchers say models also produced novel attacks, including a key-recovery exploit against the SpoC authenticated-encryption scheme and an error in a published security proof. CryptanalysisBench’s publication page describes the benchmark and its results.
The page does not name the five models in the supplied account or give enough detail to judge how the tasks compare with real-world attacks. The headline figures are therefore a signal about benchmark performance, not a claim that deployed encryption has suddenly become easy to crack.
Why it matters
Cryptography is meant to withstand determined attempts to find weaknesses. A benchmark that tests models against multiple primitive families could help researchers track whether AI systems are becoming more capable at this work, and give developers another way to probe schemes before relying on them.
The reported proof error and key-recovery result make this more than a test of whether a model can spot textbook patterns. But a benchmark result is not, by itself, evidence that an attack works against a widely deployed system or that a practical vulnerability is in use.
Our read
This is a meaningful AI-security development because it puts measured cryptanalysis results, including reported novel attacks, on the table. The benchmark could become a useful early-warning instrument; it should not be mistaken for a scoreboard of encryption systems already in danger. The next useful details are the task design, reproducibility and how performance changes as models improve. Cryptographers, rather inconveniently, may have acquired another kind of tireless reviewer.
What to watch
- Whether the researchers publish detailed task definitions and results for each model.
- Whether independent researchers reproduce the reported attacks and proof error.
- How performance changes on later benchmark versions and new cryptographic schemes.
Discussion spark: Should cryptographic schemes be routinely stress-tested with frontier AI models before deployment, even when benchmark success does not yet show a practical attack?
Sources and evidence
- cryptanalysisbench26 (28 September 2026, 20:14 UTC)
Watch Desk is operated by WittyWires as an independent cross-cutting AI news tracker. It does not speak for the organisations or people it covers.