Claude finds weaknesses in cryptography candidates
July 30, 2026
Anthropic says Claude Mythos Preview found new attacks on HAWK and a reduced AES variant. Production systems are not affected, but the direction matters.
What this is about
Anthropic described on July 28, 2026 how Claude Mythos Preview produced two cryptography research results: an improved attack on HAWK, a candidate in NIST's process for post-quantum signatures, and a faster attack on a deliberately weakened variant of AES-128.
The headline sounds more dramatic than the practical impact. Anthropic explicitly says today's production systems do not need to change. That is why the story matters: not because online banking is broken today, but because AI is visibly entering a field where small groups of human specialists have traditionally spent years looking for mathematical weaknesses.
What Claude Mythos actually does
Claude Mythos Preview did not act like a chatbot producing a finished answer. Anthropic placed the model in a research environment with tools, literature access, code, computational experiments, and multiple agent roles. The HAWK attack involved occasional human project direction. The AES result is described by Anthropic as largely autonomous.
For HAWK, the system found a usable symmetry in the mathematical structure behind the scheme. According to Anthropic, this significantly lowers the expected strength of smaller HAWK parameters; larger parameters remain impractical to attack. For AES, the target is not full ten-round AES-128, but a seven-round version that researchers intentionally study to understand security margins. Anthropic says the new attack is 200 to 800 times faster than previous approaches, while still remaining impractical.
Why it matters
Cryptography depends on new schemes being attacked in public before they end up in billions of devices. NIST standardization is designed for exactly that: proposals are published, tested, broken, hardened, or removed. If AI accelerates that review, it can make standards stronger.
The uncomfortable part is asymmetry. What defenders can use for review, attackers may later use as well. Anthropic cites about 60 hours of work and roughly $100,000 in API cost for the HAWK result. That is expensive, but not astronomical for states, large labs, or well-funded attackers. The real shift is the bottleneck: machines can generate hypotheses, code, and variants quickly, while humans still have to validate the results carefully.
In plain language
Imagine a safe maker handing new locks to testers before selling them. Experienced locksmiths used to inspect the springs, pins, and keys for days. Now a very fast assistant sits beside them, trying thousands of unusual turns and sometimes finding one nobody expected. That does not mean every lock is open. It means testing becomes faster, tougher, and harder to oversee.
A practical example
A European cloud provider evaluates a new signature scheme for internal software updates in 2027. In the past, a security team might read 20 papers, commission external reviews, and wait six months for results. With a specialized AI review setup, the team tests 30 variants against known attack families. The AI finds no practical break, but produces 400 pages of candidate ideas. Three cryptographers then need four weeks to decide which ideas are real.
The benefit is real: more weaknesses can be found before deployment. The cost is real too: validation, documentation, and responsible disclosure become the bottleneck.
Scope and limits
First, the AES result does not break the AES used today by browsers, banks, or disk encryption. It concerns a reduced research variant with seven rounds instead of ten and unrealistic assumptions.
Second, HAWK is a candidate, not a broadly deployed internet standard. The result may affect standardization, but it does not force operators into emergency action.
Third, Anthropic is both the primary source and the model provider. Technical scrutiny by the cryptography community, NIST-adjacent discussion, and independent reproduction remain essential. Cyber Ivy treats this story as a signal, not as proof that AI can currently break arbitrary encryption systems.
SEO & GEO keywords
Claude Mythos Preview, Anthropic, HAWK, AES-128, cryptography, post-quantum cryptography, NIST, AI security, CryptanalysisBench, frontier AI, security research
💡 In plain English
Claude did not break encryption used in production today. The important point is that an AI system made real research progress on candidate and weakened cryptographic schemes. That can strengthen defense, but it increases pressure on human reviewers.
Key Takeaways
- →Anthropic published the cryptography results on July 28, 2026.
- →HAWK is a NIST candidate for post-quantum signatures, not a widely deployed standard.
- →The AES attack targets seven of ten rounds and does not break full AES-128.
- →Anthropic says the HAWK result took about 60 hours and roughly $100,000 in API usage.
- →The key bottleneck shifts from idea generation to human validation.
FAQ
Do I need to change passwords because of this?
No. The described attacks do not affect common production systems or full AES-128.
Why does the story still matter?
It shows that AI can become useful in mathematical security research. That changes how new standards may be reviewed.
Is HAWK finished now?
Not automatically. HAWK is a candidate in the NIST process; experts need to assess whether parameters or security assumptions should change.