On July 28, 2026, Anthropic published a research post describing how Claude Mythos Preview discovered mathematical weaknesses in cryptographic algorithms themselves. This article summarizes what that announcement means and what it doesn't, based strictly on the official post.
🟢 Model references match the current lineup · model notice · Fable subscription
Fable 5 and 5.1 subscription (updated September 7, 2026): Claude Fable 5.1, released September 1, 2026, is the current Fable model and Fable 5 is now legacy. Plan terms are the same for both — Max and Team Premium plans include Fable at up to 50% of the weekly usage limit; Pro and Team Standard use usage credits (
What exactly did Claude do in this research?
Claude Mythos Preview found new, stronger attacks against two real cryptographic algorithms. One is an attack on HAWK, a post-quantum (designed to resist future quantum computers) digital signature scheme; the other is an attack on a weakened, round-reduced version of AES, the world's most widely used symmetric cipher. Neither result affects live systems right now, but both are cited as evidence that AI models can find flaws in the mathematical design of cryptographic algorithms, not just in how programmers implement them.
Terms explained
- Cryptography — the mathematical study of methods for keeping data private and secure
- Digital signature scheme — a cryptographic technique used to verify that a website or file is authentic and hasn't been tampered with
- Symmetric cipher — an encryption method that uses the same key to both encrypt and decrypt data, used to protect internet traffic
- NIST — the US National Institute of Standards and Technology, the US government body that sets cryptographic standards used worldwide
- Post-quantum cryptography — cryptography designed to remain secure even against future quantum computers
- HAWK — a digital signature algorithm currently a candidate in NIST's post-quantum standardization process
- AES (Advanced Encryption Standard) — the symmetric cipher adopted by NIST in 2001 and still the most widely used encryption standard today
- Round-reduced version — an intentionally weakened version of a cipher with fewer repeated encryption rounds, used for research purposes
What does the HAWK attack mean?
The key point of the HAWK result is that Claude improved an attack on an algorithm that had already survived two rounds of expert human review over two years — and did so in just 60 hours. HAWK is a third-round candidate in NIST's post-quantum cryptography competition, launched in 2022. One Anthropic researcher worked together with Mythos Preview to improve the best-known attack against HAWK, effectively cutting its key strength in half. However, HAWK is only a candidate and is not yet deployed as a standard in production systems, so this attack does not put any live service at risk today.
What does the AES attack mean?
The key point of the AES result is that, on a round-reduced (weakened) version of AES, the attack eliminated one of the guesses an attacker would otherwise need to make — improving the speed of previous best attacks by 200-800x. This work was carried out almost entirely autonomously by Mythos Preview, using a scaffold (an automated setup that lets an AI repeat and iterate on a task by itself) built by another researcher. Importantly, this attack targets a reduced-round experimental version of AES and does not break the full cipher used in real systems.
| HAWK attack | AES (round-reduced) attack | |
|---|---|---|
| Target | Post-quantum digital signature candidate | Weakened version of the most widely used symmetric cipher |
| Claude's role | Worked with a researcher; improved the attack in 60 hours | Ran almost fully autonomously on a scaffold |
| Result | Key strength effectively cut in half | Best known attack speed improved 200-800x |
| Impact on live systems | None (HAWK is not yet an adopted standard) | None (targets a reduced version, not full AES) |
Does this put the internet services I use at risk right now?
No. Anthropic explicitly states that neither result has a practical impact on today's computer systems and that no production software needs to change because of them. HAWK isn't deployed in production yet — it's still a candidate under review — and the AES attack targets a reduced-round research version, not the full cipher used in real systems. Still, Anthropic presents this as evidence that frontier AI models could help find flaws in cryptographic systems both before and after real-world deployment.
How was this disclosed responsibly?
Anthropic says it followed responsible disclosure procedures throughout this research. It worked with academics to confirm the validity of its findings, shared the HAWK attack with HAWK's authors in June, and coordinated public disclosure on NIST's public mailing list to coincide with the release of its results. Anthropic also shared advance copies with US government and industry partners and discussed the implications with them.
Can other researchers try this kind of work themselves?
Yes. Anthropic partnered with academics at ETH Zurich, Tel Aviv University, and TU Berlin to build and release CryptanalysisBench, a benchmark (a standardized test) that packages together many cryptographic ciphers so other researchers can evaluate language models' cryptanalysis capabilities.
Can I use Claude Mythos Preview myself right now?
Not as a model you can simply pick from a menu. The Mythos line used in this research is a frontier research track that Anthropic uses to test model capability at the extreme edge, and it's currently made available only to a limited set of approved organizations. It's worth knowing that this is handled separately from the Opus, Sonnet, and Haiku models generally available on Claude.ai and the API.
Why does this research matter?
This research matters because, until now, Claude has mostly been used to find "implementation bugs" — mistakes programmers make when coding an algorithm. This time, it found mathematical weaknesses in the design of the algorithms themselves, algorithms that had already been reviewed by human experts for years. Anthropic frames this as cryptography research working as intended: stress-testing algorithms to build trust and ultimately make systems more secure.
Frequently asked questions
Q. Do I need to do anything if a system I use relies on HAWK or AES?
No action is needed. HAWK is not yet an adopted standard used in production, and the AES attack targets a reduced-round research version, not the full AES cipher used in real systems.
Q. Did Claude find these results completely on its own?
It varied by result. The HAWK attack was developed with an Anthropic researcher working alongside Mythos Preview, while the AES attack was found almost fully autonomously by Mythos Preview running on a scaffold built by another researcher.
Q. How long did this take and how much did it cost?
Both results were developed over the course of about a week, and Anthropic states each cost roughly $100,000 in API costs to produce.
Q. When and how was this disclosed?
Anthropic published the findings on its official research blog on July 28, 2026, following a responsible disclosure process — sharing the HAWK attack with its authors in June, then coordinating public release with the NIST mailing list announcement.
This announcement doesn't mean Claude poses an immediate threat to the services you use — it shows AI being used as a tool to stress-test cryptographic systems more thoroughly. Results from limited-access research models like Mythos Preview are worth watching as this area develops.
Related articles
- Claude web search & Research — using live search and deep research well
- How Well Does Claude Control Robots? Anthropic's Robotics Research Explained
- Claude Projects, Artifacts & Memory: What They Are and How to Use Them
- How to Write Prompts for Claude: 6 Core Principles from the Official Docs
- How to Use Claude Projects: Knowledge & Instructions to Cut Repetition