Discovering Cryptographic Weaknesses with Claude

TL;DR Highlight

Claude Mythos Preview halved HAWK's effective key strength and sped up a reduced-round AES attack 200-800x.

Who Should Read

A backend developer designing long-running agent harnesses who wants to know how to prompt agents that give up early.

Core Mechanics

The HAWK attack halves effective key strength
- Claude Mythos Preview found a nontrivial automorphism (a symmetry) in the lattice behind HAWK, a post-quantum signature candidate.
- Full key recovery cost for HAWK-256 dropped from 2^64 to 2^38.
- Keys must be doubled in size to restore the same security level.
7-round AES attack is 200-800x faster
- AES-128 has 10 rounds, and this attack only works on a 7-round reduced version.
- A fingerprint algorithm called the Möbius Bridge removed one of the 256 guesses in a prior meet-in-the-middle attack.
- Depending on how runtime is measured, it is 200-800x faster than the previous best attack.
No impact on production systems
- HAWK is only a candidate and is not deployed.
- The AES attack does not break the full cipher.
- The post says no production software has to change.
Each result cost roughly $100,000 in API cost
- The HAWK attack took about 60 hours of work by one researcher with Claude over a week.

Community Reactions

A commenter argued this counters prompt engineering hype
- The published prompts were full of typos yet worked, so the commenter said RLHF and fine-tuning exist to make such prompts work with no fuss.
Commenters questioned the $100k opportunity cost
- One said the money might buy more long-term on a researcher or grad student, since results may not be reproducible.
- Another said at $1k it would be worth throwing at a large problem space like fuzzing.
A commenter said the HAWK framing is overstated
- They noted that an automorphism attack was already known in theory and only its accessibility was open.
- They saw it as a concrete case rather than a wholly unknown weakness.
A commenter doubted the two-agent collaboration story
- They suggested it may just reflect the stochastic nature of each thread.
- They proposed replaying the search with context perturbations to see how often it finds the path.

How to Apply

Add a goal-reminder line to your agent prompt
- If you run a long agent job today, add 'the goal is a novel result, not an easy one' to the system prompt.
- When the agent tries to change the target or quit, send one line telling it not to change the goal.
Handle 'impossible' answers with a retry loop
- Do not stop when an agent says something is impossible.
- Automatically re-prompt it to propose hypotheses and validate or refute each with experiments.
Make outputs verifiable by design
- Require the agent to ship a runnable verification script with its result.
- Automatic checks, like recovering a key, are much faster than reading and judging by hand.

Try It

copy & run

# Paste into Claude or ChatGPT (simplified version of the post's method: handling agents that give up)

You are a research agent. Your goal is to find a genuinely new approach to {{problem}} that beats existing ones.
Easy solutions or known methods are not the goal. Look for hard, publishable results.
Before concluding something is impossible, you must do the following.
1. Propose at least 3 hypotheses.
2. Validate or refute each with experiments or code.
3. Record why each was refuted and feed that into the next hypothesis.

At the end, provide a runnable script that verifies your result.

Next thing to try: if the agent tries to switch targets, reply 'do not change the target, keep working on this problem'.

Terminology

HAWKA digital signature scheme designed to stay secure against quantum computers. It is a third-round NIST standardization candidate.

reduced-round AESAES-128 normally has 10 rounds, and this is a weakened version with fewer, such as 7. Researchers use it to study attack techniques.

meet-in-the-middle attackAn attack that stores intermediate results in a big lookup table and reuses them. It trades memory for time.

scaffoldThe harness around the model that lets Claude pose hypotheses and run experiments autonomously. It is similar to an agent harness.

CryptanalysisBenchA benchmark that bundles many ciphers to test LLM cryptanalysis ability. It was built with ETH Zurich, Tel Aviv University, and TU Berlin researchers.