Measuring Human Cryptanalysis Effort

Matthew Green has a new site that tries to estimate the number of cryptanalyst-hours that have been spent analyzing different public-key algorithms:

Abstract: We estimate how much public cryptanalysis each public-key assumption has received, using the literature itself as the instrument. From roughly sixteen thousand candidate papers we classify cryptanalytic work on ECDLP, structured and plain lattices, syndrome decoding and their comparators, then convert papers into estimated research-allocation years through a person-level author map. These publication-based estimates may substantially overstate actual time spent. Lattices have overtaken elliptic curves in this cumulative measure only recently; we express that history as inference budgets under stated productivity assumptions. The study was conducted by AI models: Claude Fable 5.1 and Claude Opus 5.5 built the pipeline, ran the analysis and wrote this page, with contributions from GPT-6 Astra. Matthew Green set the questions, made the key methodological decisions and oversaw the work.

There are a lot of caveats to this analysis. It relies on published work, so anything unpublished does not count. This includes both academic cryptanalysis that didn’t result in a paper, and intelligence agency cryptanalysis done in secret. But it’s a good first stab at a metric of this sort.

Posted on October 15, 2026 at 12:24 PM • 3 Comments

Comments

Morley • October 15, 2026 1:23 PM

The work was done by AI, so I have no idea how accurate it is. Matthew needs to explain how he “oversaw” the work instead of crediting models as if they were people.

Rontea • October 15, 2026 1:57 PM

In the harmony of collective effort, the distinction between human and machine fades, and what remains is the shared aspiration toward clarity and the benefit of all beings.

Clive Robinson • October 15, 2026 3:20 PM

@ Morely, ALL,

With regards,

“The work was done by AI, so I have no idea how accurate it is.”

The same has done in the past and still applies to work done by humans now and into the future just as much.

It’s just one of the reasons we have for the “review process”. For not just academic papers but nearly all new science and engineering design work projects.

Because we accept that all “humans err” and some humans distort the truth for their own and others advantage.

Yet when we design machines to emulate human characteristics we expect them to be 100% what we want, not what humans actually are.

If the machine is fully deterministic and we draw up the operational rules correctly then we stand a chance of getting 100% what we want. However only “a chance” as seen by software development and some engineering projects, things fail for reasons not understood before and in some cases after first use / testing, and when “noise” enters the processing system.

Worse we have since the work of Claude Shannon realised that we can use stochastic / random processes to advantage as they can save time or other resources.

Which unfortunately,

1, Increases complexity
2, Breaks deterministic behaviour

Both of which make any list of operational rules effectively incomplete.

Thus errant / unexpected behaviour should be expected as a matter of normal usage.

Which means that the question arises as to,

“What can be done to minimise the errant behaviour or recognise incorrect output issues.”

This has been an issue under consideration for around a century and a half.

We have in that time developed some partial solutions but none of them we now know can be 100%.

Thus there is a penalty for using stochastic or random processes which is uncertainty in the results. It’s something that can not be avoided though there are tricks to minimize the uncertainty by using “feedback techniques”.

One of the issues with LLMs is that they can be seen as “Digital Signal Processing”(DSP) networks in “Feed Forward”(FF) mode. With The problem of FF mode being is that noise, stochastic, or random, processes get amplified.

The usual way to deal with this is to do the process multiple times in parallel then take the arithmetic mean of the results. This obviously rapidly increases resource usage issues, which is why such methods are mostly not used[1].

[1] The main exception for this is where “Feed Back”(FB) methods are not realistically possible. One such is for error correction on long distance or otherwise constrained communications paths or with recorded media. Which is why we have “Forward Error Correction”(FEC) techniques and systems taking considerably extra resources.

Leave a comment

Blog moderation policy

Login

Allowed HTML <a href="URL"> • <em> <cite> <i> • <strong> <b> • <sub> <sup> • <ul> <ol> <li> • <blockquote> <pre> Markdown Extra syntax via https://michelf.ca/projects/php-markdown/extra/

Sidebar photo of Bruce Schneier by Joe MacInnis.