July 29, 2026
Locked out? Logged on.
Some thoughts about Anthropic's new cryptanalysis results
AI pokes a future lock, barely dents today’s vault, and the comments go feral
TLDR: Anthropic’s AI found a meaningful weakness in a future encryption candidate, but its splashier AES result only affects a weakened practice version, not the locks protecting daily life. Commenters turned it into a brawl over AI hype, hidden model access, and whether this proves the bots are way past “autocomplete.”
Anthropic dropped two shiny AI-made codebreaking results, and the internet instantly split into panic mode, eye-roll mode, and this is kind of hilarious mode. The bigger deal is a hit on HAWK, a not-yet-adopted “future-proof” digital signature idea meant to survive quantum computers. Translation: it’s not protecting your bank account today, but it was a serious candidate for tomorrow. The spicier part, as commenters kept repeating, is that the AI didn’t discover magic new math — it basically took a pile of known tricks and worked harder than the humans had. One commenter summed up the vibe brutally: Anthropic seems to have just told the model to go get results and “strapped its nose to the grindstone until it found some.”
The AES side of the story — the one with the scary name ordinary people might recognize because AES is the encryption used all over modern tech — sparked way more skepticism. Yes, Anthropic improved an attack on a weakened test version, but commenters were quick to dunk on any “AI breaks the internet” narrative. This wasn’t the full-strength lock, and several people treated that part like trailer hype for a movie that never delivers.
Then came the meta-drama. One reader mocked the article’s avalanche of AI-generated summaries with a snarky “did I just read two summaries/TLDRs in a row,” while another argued the real bombshell is that anyone still calling these models “glorified autocomplete” needs to stop immediately. Bonus intrigue: Simon Willison stirred the pot by claiming Mythos is basically already out — just hidden behind partner access and filters. So the comment section’s verdict was deliciously messy: interesting result, limited immediate danger, and absolutely maximum discourse.
Key Points
- •Anthropic reported two Claude Mythos cryptanalysis results: an attack on HAWK and an improved attack on reduced-round AES.
- •The HAWK result targets a proposed, non-deployed post-quantum signature scheme and does not transfer to the Falcon signature scheme.
- •The HAWK attack remains exponential-time but is described as roughly halving the scheme’s security, potentially requiring larger keys to compensate.
- •Anthropic demonstrated the HAWK attack with working code on a weakened challenge instance rather than proposed real-deployment parameters.
- •The AES result applies to a 7-round reduced-round version of AES, and the article says such attacks are already known and that this new result is only a modest improvement.