Anthropic Says Claude Mythos Has Discovered Weaknesses In Cryptographic Algorithms That Keep Data Safe

Anthropic's unreleased Claude Mythos Preview model has discovered two cryptanalytic weaknesses: an attack that halves the effective key strength of the post-quantum digital signature scheme HAWK, and a new attack technique on a reduced-round version of AES. Anthropic says neither result threatens any system in production today, but both demonstrate AI's capability to perform research-level cryptanalysis.

AI isn’t just finding gaps in cybersecurity — it’s now beginning to find weaknesses in the encryption methods that underpin data security on the internet. Anthropic has revealed that its unreleased Claude Mythos Preview https://officechai.com/ai/claude-mythos-preview-benchmarks-swe-bench-pro/ model, developed as part of the company’s Project Glasswing cybersecurity initiative, was used to discover two significant cryptanalytic results: an attack that halves the effective key strength of a post-quantum digital signature scheme called HAWK, and a new attack technique on a reduced-round version of AES, the encryption standard used to secure most of the world’s internet traffic. Anthropic says neither result threatens any system in production today, but both point to AI models being capable of research-level work in a field that has traditionally depended on years of expert scrutiny. What these algorithms actually do To understand why this matters, it helps to know what these two systems are doing every time someone loads a website or sends an email. When a browser connects to a site, it first needs to confirm it’s actually talking to the real server and not an impostor. That job falls to a digital signature scheme, a mathematical method that lets a server prove its identity without revealing anything an attacker could steal and reuse. HAWK is one of several candidate schemes competing in a multi-year process run by the US National Institute of Standards and Technology NIST to pick new signature systems that would remain secure even against a future quantum computer, since machines of that kind could eventually break signature schemes like RSA that the internet has relied on for decades. Once identity is confirmed, the actual data — banking details, messages, files — gets scrambled using a symmetric cipher, where both sides share the same secret key to encrypt and decrypt information. AES, adopted by NIST back in 2001, is the cipher doing this job in the overwhelming majority of cases, from HTTPS traffic to Wi-Fi to disk encryption. It works by running data through the same scrambling process repeatedly, in AES-128’s case ten times over, called rounds. What Anthropic found The HAWK attack came out of a week-long collaboration between one Anthropic researcher and Mythos Preview, working inside a multi-agent setup where several instances of the model operated together, occasionally disagreeing with each other before converging on an idea. Mythos identified a mathematical symmetry in HAWK’s underlying lattice structure that nobody had previously exploited, one that lets an attacker search for the secret key far faster than expected. In concrete terms, breaking the smaller HAWK-256 variant was believed to cost on the order of 2^64 operations; Mythos showed it can be done in roughly 2^38. That’s still an enormous number of operations and no practical threat today, but it means HAWK’s proposed key sizes would need to double to restore the intended security margin, which undercuts much of what made the scheme attractive as a candidate in the first place. The AES result took a different shape entirely. A separate researcher built an agentic scaffold and simply asked Mythos to find something better than the best published attack on 7-round AES-128, a heavily studied academic sub-problem used to gauge how much safety margin the full 10-round cipher carries. The model’s early runs were, by Anthropic’s own account, unproductive — Mythos repeatedly insisted the problem was already solved as well as it could be, telling researchers there was “nothing easy to find” in what it called the most-studied block cipher in existence. It took a handful of blunt human nudges, including messages telling the model to stop looking for easy wins and act like a serious researcher, before Mythos began producing anything new. Over the following three days, largely without further human input, Mythos generated hundreds of millions of tokens of research and eventually landed on an idea it named the Möbius Bridge, a new way of fingerprinting intermediate values in the attack that eliminates a step requiring 2^56 guesses altogether. The result is an attack between 200 and 800 times faster than the previous best-known method against 7-round AES. It does not come close to breaking full AES-128, and Anthropic notes that even the underlying academic attack this builds on would cost hundreds of millions of dollars to actually carry out. Verifying that Mythos’s claim was even correct took Anthropic researchers, who by their own admission aren’t cryptography experts, nearly a month of additional checking. Why Anthropic is talking about this now Anthropic frames both findings as cryptography working the way it’s supposed to: schemes get proposed, picked apart, and either strengthened or discarded before they’re trusted with real data. That’s explicitly the point of NIST’s public standardization process, and the company notes that a late-stage flaw isn’t unprecedented — one earlier NIST post-quantum candidate, SIKE, was completely broken on a laptop within an hour once the right technique was found. What’s changed is who’s doing the finding. Anthropic says it briefed the HAWK authors ahead of time and coordinated public disclosure through NIST’s official mailing list, and that it consulted government and industry partners before publishing either result. The company has also built a benchmark called CryptanalysisBench with academics from ETH Zurich, Tel Aviv University, and the University of Haifa specifically to track how fast this capability is moving across different AI models over time. Anthropic’s broader argument is that this won’t stay confined to two isolated results. The company says it has also used Mythos to develop a practical, end-to-end working attack on a 13-round version of the LEA cipher, a lightweight standard used in low-power devices, and a full key-recovery attack on 6 rounds of the Serpent-128 cipher, along with smaller improvements against Salsa20, the Poseidon hash function, and SHA-1. None of these break the full ciphers in use today. But Anthropic’s pitch is that many cryptographic systems in circulation have had far less expert attention than AES or the NIST finalists, and that the same capability now being used to find these weaknesses can just as usefully be turned toward designing stronger algorithms in the first place. The company, whose Opus 5 model launched days ago https://officechai.com/ai/claude-opus-5-benchmarks/ with cybersecurity vulnerability-scanning restrictions of its own, says it plans to host an academic workshop in the coming weeks to work through how the security research community should respond as language models start producing results that outpace the human bandwidth needed to verify them.