Does this break AES in production?

No. The Möbius Bridge attack targets 7-round reduced AES, not the full 10-round AES used in production. Anthropic states clearly that no deployed cryptosystem is currently affected. The result is a cryptanalytic advance on a weakened variant, useful for stress-testing where the security margins really lie.

What happens to HAWK now?

HAWK is a NIST post-quantum signature candidate, not yet standardized or deployed. Claude Mythos found a lattice automorphism that roughly halves the effective key strength, so HAWK now needs larger parameters to hit its target security level. The finding is being fed into NIST's review of the scheme.

How autonomous was the AI in this work?

The two results sit at different points on the autonomy scale. HAWK took about 60 hours of tight researcher collaboration with Claude Mythos and around $100,000 in API costs. The AES Möbius Bridge attack ran near-autonomously for three days, producing billions of tokens with only three brief human encouragement prompts.

Is the code released?

Yes, Anthropic published a demo repository at github.com/anthropics/cryptography-research-demo under Apache-2.0, containing the AES, HAWK and LEA cryptanalysis code. The repo is marked as an unmaintained research artifact rather than a supported library.

What is CryptanalysisBench?

CryptanalysisBench (arXiv:2607.18538, submitted July 20, 2026) is a companion benchmark of 191 tasks across six families of cryptographic primitives. It was built to track whether frontier AI models can perform real cryptanalysis, with results showing they now break between 65-86% of simpler schemes and produce genuinely novel attacks on harder ones.

Anthropic · 2026-07-28 · major

Anthropic uses Claude Mythos to weaken HAWK and 7-round AES — Apache-2.0 demo code released

Anthropic's cryptographic-weaknesses research shows Claude Mythos Preview finding a lattice automorphism that halves HAWK's post-quantum key strength and a new Möbius Bridge attack that speeds 7-round AES cryptanalysis 200-800x, all with Apache-2.0 demo code.

Anthropic research illustration for discovering cryptographic weaknesses with Claude

Claude Mythos Preview found new mathematical attacks that halve HAWK's post-quantum key strength and speed 7-round AES cryptanalysis 200-800x, released with open demo code.

Key specs

Hawk key strength reduction	≈50%
7 round aes speedup	200-800×
Cryptanalysis bench tasks	191
Demo code license	Apache-2.0

Quick facts

Maker	Anthropic + academic collaborators
Model	Claude Mythos Preview
Main results	HAWK lattice automorphism; 7-round AES Möbius Bridge attack
Extra findings	Attacks on 13-round LEA and 6-round Serpent
Code	github.com/anthropics/cryptography-research-demo (Apache-2.0)
Benchmark	CryptanalysisBench (arXiv 2607.18538, 191 tasks)
Production impact	None — attacks target non-deployed HAWK candidate and reduced-round AES

What is it?

The research post describes two new cryptanalytic results from Claude Mythos Preview, Anthropic's restricted security model. On HAWK, a NIST post-quantum signature candidate, Claude located a nontrivial automorphism in the underlying lattice that theorists suspected existed but had never found, effectively halving key strength. On 7-round reduced AES, Claude produced a new attack the team calls the Möbius Bridge fingerprint.

How does it work?

The HAWK result came from about 60 hours of researcher-guided work with Claude Mythos, at roughly $100,000 in API costs. The AES Möbius Bridge attack ran near-autonomously over three days, producing billions of tokens with only three brief prompts from the human. Both were validated by human cryptographers and published as PDFs alongside a companion CryptanalysisBench paper (arXiv:2607.18538) that scores five frontier models on 191 cryptanalytic tasks.

Why does it matter?

Cryptanalysis is one of the highest-stakes places to test model capability: real progress means real weakening of algorithms defenders depend on. The HAWK and AES results are theoretical — HAWK is not deployed, and the AES attack is on 7 of 10 rounds — so no production system is at risk today. But shipping open demo code plus a benchmark means the wider cryptography community can now rerun, extend and stress-test what Claude found.

Who is it for?

cryptographers, post-quantum standards reviewers, AI-safety researchers

Frequently asked questions

Does this break AES in production?: No. The Möbius Bridge attack targets 7-round reduced AES, not the full 10-round AES used in production. Anthropic states clearly that no deployed cryptosystem is currently affected. The result is a cryptanalytic advance on a weakened variant, useful for stress-testing where the security margins really lie.
What happens to HAWK now?: HAWK is a NIST post-quantum signature candidate, not yet standardized or deployed. Claude Mythos found a lattice automorphism that roughly halves the effective key strength, so HAWK now needs larger parameters to hit its target security level. The finding is being fed into NIST's review of the scheme.
How autonomous was the AI in this work?: The two results sit at different points on the autonomy scale. HAWK took about 60 hours of tight researcher collaboration with Claude Mythos and around $100,000 in API costs. The AES Möbius Bridge attack ran near-autonomously for three days, producing billions of tokens with only three brief human encouragement prompts.
Is the code released?: Yes, Anthropic published a demo repository at github.com/anthropics/cryptography-research-demo under Apache-2.0, containing the AES, HAWK and LEA cryptanalysis code. The repo is marked as an unmaintained research artifact rather than a supported library.
What is CryptanalysisBench?: CryptanalysisBench (arXiv:2607.18538, submitted July 20, 2026) is a companion benchmark of 191 tasks across six families of cryptographic primitives. It was built to track whether frontier AI models can perform real cryptanalysis, with results showing they now break between 65-86% of simpler schemes and produce genuinely novel attacks on harder ones.

Try it

git clone https://github.com/anthropics/cryptography-research-demo