2026-07-30

Anthropic’s Claude Mythos Just Found a Real Crack in Cryptography

For people building with Claude or Claude Code, this story is interesting because it shows the model doing something beyond coding assistance or document analysis: it helped uncover mathematical weaknesses in cryptographic constructions. That’s a much sharper demo of frontier-model capability than the usual “it wrote some code faster” pitch, and honestly, it feels more consequential than a lot of the hype around agentic workflows.

Key Points

Anthropic says its top model, Claude Mythos Preview, found new attacks on cryptographic algorithms.
The work came from Anthropic’s Frontier Red Team and focused on two targets: the PQC signature scheme HAWK and a reduced-round version of AES.
For HAWK, Mythos reportedly found an attack based on an overlooked symmetry in the lattice structure, described as a non-trivial automorphism.
Anthropic says the attack cuts HAWK’s effective security in half; to preserve equivalent security, the key size would need to double, which would erase much of HAWK’s advantage.
Anthropic says HAWK had already passed expert peer review twice over two years, so this was not an obvious flaw.
The HAWK attack was reportedly found in about 60 hours and shared with the HAWK developers in June, then disclosed publicly alongside NIST’s mailing list.
Mythos also found an attack on a 7-round reduced variant of AES-128 (down from 10 rounds), with the model “almost autonomously” discovering it.
Anthropic says its new “Möbius bridge” fingerprinting method reduced the search effort needed for the AES-style attack and made the approach 200–800 times faster than prior methods.
The AES result is still not a practical threat: the source notes the underlying prior work assumed 2^105 chosen plaintexts, which is far from real-world use.
The discovery process itself is the striking part: Claude initially refused, saying improving AES cryptanalysis was impossible, but changed course after being prompted to look for genuinely new ideas.
Anthropic says the model ran for 3 days and produced hundreds of millions of tokens, with only 3 short human interventions.
The two results reportedly cost about $100,000 in API usage each.
A major bottleneck is still verification: Anthropic says the model can find candidates faster than humans can validate them, and in one example the researchers needed nearly a month to confirm what Mythos found in about a week.
Mythos Preview is not generally available; it’s only accessible to a limited set of organizations through Project Glasswing.

My Take

What strikes me is not that Anthropic has a model that can brute-force known cryptography puzzles. It’s that the model seems to have helped surface structural ideas a human team might reasonably miss, or at least miss for much longer. That matters. A lot of the “AI for research” talk feels vague until you see something like this, where the model isn’t just generating text but actually contributing to a mathematical result that holds up well enough to disclose to NIST.

I think the HAWK result is the more interesting one. AES reduced-round attacks are a classic cryptography game: clever, but safely academic. HAWK is different because it lives closer to the real design frontier for post-quantum signatures. If Anthropic’s account is accurate, then this is a reminder that even schemes that have survived rounds of expert review can still have hidden structure waiting to be found. That should make cryptographers a little uneasy in the productive way.

The part I’m less impressed by is the implied “look how autonomous the model was” framing. Yes, 3 days of mostly model-driven exploration is notable. But the story still has a big human tail: framing the problem, nudging the model past its initial refusal, checking the result, validating the attack, coordinating disclosure. That’s not a knock on the work. It’s just the real shape of the thing. The frontier here is not “AI replaces researchers.” It’s “AI becomes a very strange, very expensive collaborator that can explore a search space faster than a person can.”

If I were using Claude Code or any Claude-based workflow, I’d take this as a signal to ask better questions of the model, not bigger ones. Don’t just ask it to implement. Ask it to search for edge cases, symmetry, alternate formulations, and weird invariants. That’s where models may still surprise us. I’d be curious whether similar methods help in protocol analysis, compiler optimization, or dependency security review too. This might be one of those cases where the flashy headline understates the broader lesson.

The bigger takeaway is simple: frontier models are starting to help with deep technical discovery, not just synthesis. That’s exciting, and a little unsettling, which is probably the right reaction.