Claude Mythos Preview finds flaws in reduced-round AES cipher
Anthropic says its Claude Mythos Preview found weaknesses in a weakened AES variant, renewing debate over AI cryptanalysis.
In Brief
- Anthropic says Claude Mythos Preview discovered weaknesses in a trimmed-down version of the AES block cipher.
- The findings concern a reduced-round AES variant, not the full standard in production use today.
- Researchers say the result is a milestone for AI-assisted cryptanalysis but not an immediate threat to real-world encryption.
Anthropic’s Claude Mythos Preview model was able to identify structural weaknesses in a weakened, reduced-round version of the AES encryption standard, the company said, marking one of the clearest signs yet that large language models can contribute to cryptanalysis. The disclosure, carried by The New York Times and surfaced on TechMeme and Hacker News, has reignited discussion about where AI’s limits in mathematical reasoning actually lie. According to The New York Times,
The result does not break the AES algorithm that protects everything from messaging apps to disk encryption. Anthropic stressed that the model succeeded on a deliberately simplified variant with fewer rounds than the real standard, a controlled setting researchers use to probe how attacks scale. Even so, the company framed the demonstration as evidence that frontier models are beginning to assist with the kind of structured reasoning that cryptography demands.
The work arrives as labs race to show their models can tackle harder technical problems. On Hacker News, security engineers were quick to separate the headline from the substance: a reduced-round cipher is a well-known testbed, not production infrastructure. Still, the demonstration matters because it shows models moving from pattern completion toward genuine analytic steps in a domain where small errors are catastrophic.
What Claude Mythos Preview actually found
Anthropic reported that Claude Mythos Preview located weaknesses in a reduced-round implementation of AES, the Advanced Encryption Standard adopted by the U.S. government in 2001 and used worldwide. Reduced-round versions strip out iterations of the cipher’s mixing process, making them tractable for analysis while preserving its structure. The model’s success on that simplified target is what the company is highlighting.
Cryptographers have long used reduced-round ciphers as a yardstick. A model that can find openings there is demonstrating reasoning transferable, in principle, to harder problems — though no one is claiming the full 10- or 14-round AES is at risk. The milestone is about trajectory, not a breach.
Anthropic emphasized caution in its messaging, and outside cryptographers have urged against overreading a single vendor claim. The demonstration is consistent with a broader push by labs to show models doing real scientific and mathematical work rather than fluent summarization.
Why the encryption result matters for AI security
For defenders, the episode is a useful stress test of how quickly automated analysis could accelerate. If models can reliably find flaws in simplified ciphers, the same tooling could eventually help audit real protocols — or, in adversarial hands, probe them. That dual-use nature is precisely why Anthropic framed the work as both a capability and a reason for careful evaluation.
The broader signal is that ‘frontier models can do math’ is no longer abstract. Whether the leap from reduced-round AES to production cryptography is years or never remains an open question, but the direction of travel is now visible. Researchers on Hacker News largely agreed the demo is real progress even if the practical risk today is negligible.
Anthropic did not immediately respond to a request for the underlying methodology, and outside cryptographers have cautioned against overreading a single vendor claim. The consensus: encouraging, interesting, and not a reason to change your password manager.
Related coverage: Amodei: Anthropic never backed an open-weights model ban · Shared Claude chats surfaced on Google
FAQ
Did Claude break real AES encryption?
No. Anthropic says Claude Mythos Preview found flaws in a reduced-round, weakened version of AES used as a research testbed, not the full standard protecting real systems.
Should I be worried about my encrypted data?
Not from this result. The demonstration targeted a simplified cipher variant; production AES with its full round count remains unbroken.
Why does this matter if it is not a real break?
It shows frontier models can contribute to structured cryptanalysis, a milestone for AI-assisted reasoning even if practical risk today is minimal.
