Anthropic researchers have used Claude Mythos Preview to discover new mathematical attacks against two widely studied encryption algorithms, achieving in roughly 60 hours what human cryptographers have not found despite years of expert review. The results, published on Anthropic’s research blog on July 28, include an attack that halves the effective key strength of HAWK, a post-quantum digital signature scheme under consideration by NIST, and a new technique that speeds up attacks on reduced-round AES by 200 to 800 times.
What is encryption? When you visit a website, your browser uses two types of mathematical tools to keep your data safe. A digital signature scheme confirms the site is genuine, and a symmetric cipher scrambles the actual data traveling between you and the server. HAWK is a candidate for the first job, designed to survive quantum computers. AES, adopted by NIST in 2001, handles the second and protects everything from HTTPS traffic to Wi-Fi passwords to disk encryption. Flaws in these systems could put billions of users’ data at risk.
🔍 THE BOTTOM LINE
An AI model has, for the first time, found mathematical flaws in cryptographic algorithms that survived years of human expert review. Neither result breaks any system in production today, but the pace of discovery suggests AI will become a standard tool for stress-testing the codes that guard the internet, and possibly for breaking ones that have received less scrutiny.
How Claude Broke HAWK
HAWK is one of the remaining third-round candidates in NIST’s multi-year process to select post-quantum digital signature schemes, systems designed to remain secure even against future quantum computers. Its security relies on the hardness of a mathematical problem called the Lattice Isomorphism Problem.
Working semi-autonomously in an agentic harness, Mythos Preview found a previously unexploited symmetry in HAWK’s lattice structure called a nontrivial automorphism. This allows a faster enumeration attack that, while still exponential, effectively cuts HAWK’s key strength in half. The smaller HAWK-256 variant was believed to cost roughly 2^64 operations to break; Mythos showed it can be done in approximately 2^38.
According to Anthropic’s research post, the discovery process involved multiple worker agents collaborating in a sandboxed environment. The key idea emerged when two agents investigated the same approach: one prematurely rejected it as infeasible, but the other found a way to exploit it fully. Finding, developing, and verifying the attack took about 60 hours, at an estimated API cost of $100,000.
Doubling HAWK’s key size to restore the intended security margin would undercut many of the reasons the scheme was considered attractive in the first place. Anthropic shared the attack with HAWK’s authors in June and coordinated disclosure through NIST’s public mailing list.
The Möbius Bridge: Attacking AES
The second result concerns AES-128, the most widely used symmetric cipher in the world. Academics regularly study weakened versions of AES, removing rounds from the full 10-round cipher to understand its security margins. Mythos was asked to improve on the best-known attack against 7-round AES.
Initially, the model refused to engage. As OfficeChai reported, Mythos repeatedly told researchers that AES was “the most-studied block cipher in existence” and that there was “nothing easy to find.” It took several blunt human nudges, including messages telling the model to stop looking for easy wins and act like a serious researcher, before Claude began producing novel ideas.
Over three days, largely without human input, Mythos generated hundreds of millions of tokens and eventually discovered what it called the Möbius Bridge, a new fingerprinting algorithm that eliminates a step requiring 256 guesses. The result is an attack between 200 and 800 times faster than the previous best-known method against 7-round AES. The attack does not threaten full AES-128, and even the underlying academic attack would cost hundreds of millions of dollars to execute.
Verifying that Mythos’s claim was correct took two Anthropic researchers, who by their own admission are not cryptography experts, nearly a month of additional checking.
What This Means for Security
Neither attack affects any production system. HAWK is only a candidate, not deployed. The AES attack works on a reduced-round variant, not the full cipher. Anthropic is explicit about this.
But the broader signal is significant. As we noted in our coverage of the Mythos vulnerability disclosure spike, AI models are already finding implementation bugs in software faster than humans can patch them. Now they are finding mathematical flaws in the algorithms themselves, which is a different category of discovery.
Anthropic says it has also used Mythos to develop a practical end-to-end attack on 13 rounds of the LEA cipher, a lightweight standard used in low-power devices, and a full key-recovery attack on 6 rounds of Serpent-128. The company has built a benchmark called CryptanalysisBench with academics at ETH Zurich, Tel Aviv University, and the University of Haifa to track how fast this capability is moving across different AI models.
The Verification Bottleneck
The limiting factor is no longer AI’s ability to find results. It is human capacity to verify them. It took one week for Mythos to discover the improved AES attack, but two researchers spent nearly a month validating it. Anthropic says the vast majority of its time over the past few months has been spent verifying Claude’s results, not generating them.
This points to a structural shift in security research. If AI models can produce novel cryptographic research autonomously, the bottleneck moves from discovery to validation. The cybersecurity community is already grappling with AI-discovered bugs overwhelming standard triage and remediation processes, as we explored in our coverage of the Mythos patching crisis. The same pattern may now apply to academic cryptography.
NZ Angle
New Zealand organisations rely on the same cryptographic standards as everyone else. Government agencies, banks, and internet providers all depend on AES for data protection and on NIST-standardised schemes for digital signatures. If AI-accelerated cryptanalysis shortens the lifespan of cryptographic algorithms, the post-quantum cryptography migration that NZ institutions are beginning to plan for could arrive sooner than expected. The Government Communications Security Bureau (GCSB), which advises NZ agencies on cryptographic standards, will be watching NIST’s post-quantum standardisation process closely.
❓ FAQ
Does this mean my online banking is at risk? No. Both attacks target non-production systems. HAWK is a candidate scheme, not deployed. The AES attack works on a reduced-round variant, not the full 10-round cipher used in production.
How much did it cost to find these flaws? Anthropic estimates each attack cost roughly $100,000 in API compute. The HAWK attack took about 60 hours of work; the AES attack took several days and generated over one billion output tokens.
What is CryptanalysisBench? A benchmark Anthropic built with academics at ETH Zurich, Tel Aviv University, and the University of Haifa to evaluate how well language models perform cryptanalysis. It packages together many cryptographic ciphers so others can test their models.
Could other AI models do this? Anthropic notes that other labs have used AI for mathematical research, including Google’s work on Erdős problems and OpenAI’s resolution of the unit distance conjecture. The CryptanalysisBench is designed to track this capability across models over time.
What should happen next? Anthropic plans to host an academic workshop to discuss how the security research community should respond as language models start producing results faster than humans can verify them. The company argues that AI should ultimately help design stronger algorithms, not just break existing ones.
🔍 THE BOTTOM LINE
An AI model found mathematical flaws in encryption algorithms that human experts had scrutinised for years. The attacks do not break any system in use today. But the discovery cost $100,000 and 60 hours of compute, compared to the years of academic effort that failed to find the same flaws. The bottleneck is shifting from discovery to verification, and the cryptography community is not yet built for a world where AI produces research faster than humans can validate it.
📰 Sources
- Anthropic Research Blog — “Discovering cryptographic weaknesses with Claude”
- OfficeChai — “Anthropic Says Claude Mythos Has Discovered Weaknesses In Cryptographic Algorithms”
- New York Times — “An Anthropic Claude AI Model Finds Flaws in Tough-to-Crack Encryption Algorithms”
- Hacker News discussion (49082022, 49085064)
