August 5, 2026
Crypto panic? Not so fast
LLMs won't break symmetric crypto
AI scared one crypto contender, but commenters say your passwords aren’t doomed yet
TLDR: Anthropic’s AI found some niche crypto weaknesses, but nothing that breaks the encryption protecting everyday systems. Commenters split between eye-rolling at the hype and warning that AI could still become dangerous by automating the tedious work humans use to find cracks.
The big headline is not that artificial intelligence just smashed the locks on the internet. It didn’t. Anthropic says its model Claude Mythos helped find fresh weaknesses in a post-quantum signature candidate called HAWK and a limited, reduced version of the famous encryption system AES. That sounds terrifying until the community rushed in with the digital equivalent of: everybody calm down. The main mood was that this is interesting research, not an apocalypse. Full-strength modern encryption still looks safe, and even the article’s author flatly says they’re not worried about AI breaking AES, ChaCha, SHA-3, or BLAKE3 anytime soon.
But the comments? Oh, they brought the drama. One camp basically said, “Nice try, but a text-predicting chatbot is not about to discover magic math and crack giant secret numbers.” Another camp was much less chill, arguing that the usual “it’s too hard” defense sounds weak because grinding through endless tests is exactly the sort of boring superhuman labor AI might be good at. That sparked the classic tech-thread split: skeptics calling hype, worriers warning that today’s toy attacks could become tomorrow’s nasty surprise.
There was also meme energy. One commenter joked that this means Silicon Valley won’t become reality after all, while another framed AI’s real danger as finding bugs in software and proofs rather than performing some movie-style crypto death blow. In other words: the math may be safe, but the comment section is absolutely under attack.
Key Points
- •Anthropic said Claude Mythos helped discover a key-recovery attack on the post-quantum signature candidate HAWK, reducing the estimated security of HAWK-512 below its 128-bit target.
- •Anthropic also reported an improved key-recovery attack on 7-round AES-128, but the article states the result is not practical and does not threaten full 10-round AES-128.
- •The article highlights Anthropic’s view that LLM-assisted cryptography research is useful, especially for formalizing cryptanalytic techniques and reasoning about difficult attacks.
- •The author says LLMs are likely to be more useful for finding errors in complexity estimates and security proofs than for breaking established cryptographic schemes.
- •Anthropic helped create CryptanalysisBench, a benchmark covering AES, ChaCha, BLAKE, and post-quantum schemes, while the article argues LLMs are unlikely to break established symmetric cryptography such as AES, ChaCha, SHA-3, and BLAKE3.