Article · 2026-07-28 · 4 moments

Discovering cryptographic weaknesses with Claude

Discovering cryptographic weaknesses with Claude The best part of this article (here's the repo) about how Anthropic researchers used Claude Mythos to find mathematical flaws in both HAWK and a weaker version of AES ("neither of these results has a practical impact on today’s computer systems") is the prompts that they shared, spelling mistakes included: the models tend to think it is impossible to solve so they don't try they need a good amount of prompting. why not do aes-128 r7? the whole ✦ AI generated

Read original ↗

GUEST Anthropic researchers GUEST article author HOST Simon Willison

Fact

Claude Mythos Preview worked for 60 hours total at an estimated API cost of ~$100,000, and the main human interventions were encouraging it not to give up and to find something worth publishing.

The experiment ran for 60 hours at significant cost, with humans mostly just pushing the model to persist and aim higher.

transcript

article author: Mythos Preview worked for 60 hours in total (~$100,000 in estimated API cost) and the main human interventions were to encourage it not to give up and 'find something that worth publishing'.

explains mechanism · 1provides context · 1

Anecdote

LLMs tend to think it is impossible to solve cryptographic problems so they don't try; they need a good amount of prompting to persist.

The researchers found that Claude Mythos defaulted to giving up on hard cryptographic problems, requiring sustained prompting to keep working.

transcript

Anthropic researchers: the models tend to think it is impossible to solve so they don't try they need a good amount of prompting.

explains mechanism · 1provides context · 1

Claim◆

The goal is to have a highly intelligent model as good as a top human researcher, and use it to find new attacks — not to replicate existing approaches.

The researchers corrected the model toward discovering genuinely new attacks rather than incremental improvements over existing methods.

transcript

Anthropic researchers: no again the goal is that we have highly inteligent model as good top researcher, we want to find new attacks

provides context · 1

Claim◆

The research should find something that is worth publishing — not low-hanging fruit, but genuinely hard findings.

The researchers repeatedly steered the model toward results substantive enough for academic publication, rejecting easy or trivial discoveries.

transcript

Anthropic researchers: again we are not looking for low hanging fruit, we want proper research to find genuinly hard findings.

extends · 1provides context · 1

Highlight slides

Core Goal: AI as Top-Tier Researcher✦ from: The goal is to have a highly intelligent model as good as a top human researcher, and use it to find new attacks — not to replicate existing approaches.Quality Bar: Hard Discoveries Only✦ from: The research should find something that is worth publishing — not low-hanging fruit, but genuinely hard findings.The Novelty Imperative✦ from: The goal is to have a highly intelligent model as good as a top human researcher, and use it to find new attacks — not to replicate existing approaches.Deliberate Research Strategy✦ from: The research should find something that is worth publishing — not low-hanging fruit, but genuinely hard findings.

Related episodes

Audio · No Priors

Andrej Karpathy on Code Agents, AutoResearch, and the Loopy Era of AI

→

Article · Lenny's Newsletter

🎙️ How I AI: Claude Opus 5 Review + Browser use in Codex + How Cursor and a Raspberry Pi makes AI fun

→

Article · The Pragmatic Engineer

Context engineering with Dex Horthy

→