Claude Did a Week of Cryptanalysis; Humans Took Nearly a Month to Confirm It Was Right
On July 28, Anthropic published a set of hard results: Claude Mythos Preview helped researchers improve a key-recovery attack on the HAWK post-quantum signature scheme and found a new attack on 7-round AES-128.
Let's set the boundaries first. It did not break the full AES used in today's production systems, nor did it make internet encryption suddenly obsolete. What's truly worth watching is that, for the first time, the speed of research and the speed of verification have diverged this far.
HAWK: from 2^64 down to 2^38
HAWK is a post-quantum digital signature candidate. Claude found a non-trivial automorphism in the lattice structure it uses, which speeds up an enumeration-based key-recovery attack.
For HAWK-256, the expected attack cost drops from about 2^64 to 2^38. That's not a minor tweak—the effective security level was nearly cut in half.
That said, the attack is still exponential-time, does not affect other NIST post-quantum signature candidates, and certainly doesn't overturn lattice cryptography as a whole. HAWK isn't deployed in production systems yet; research at this stage is supposed to dig out pitfalls before standardization.
The research process itself was interesting: multiple worker agents collaborated in sandboxes with access to Python, Sage, and cryptography literature. One worker judged the key idea infeasible and gave up too early; another kept pursuing it and found the full exploitation path. Only after exchanging notes did the two converge. About 60 hours.
This is not truth automatically emerging from copying one model ten times. Disagreement, record-keeping, re-checking, and mechanisms to keep exploring are what actually make multi-agent setups productive.
AES: Möbius Bridge compresses one stage of work 256-fold
The second study targets 7-round AES-128. Full AES-128 has 10 rounds, so this is a classic reduced-round research target in cryptanalysis.
Claude proposed a fingerprinting method called Möbius Bridge, replacing a stage that originally required enumerating 256 values with a guess-invariant fingerprint. That stage's workload drops 256-fold; combined with other optimizations, the overall attack is roughly 200 to 800 times faster than the previous best.
The numbers are striking, but real-world threat is very low: the attack requires about 2^105 chosen plaintexts—completely impractical—and doesn't apply to full 10-round AES.
The model initially refused to work on this, reasoning that AES analysis is mature and progress would be hard. Researchers kept correcting it, telling it not to chase low-hanging fruit. Within three days the model produced hundreds of millions of tokens; results formed in about a week, with nearly a billion tokens consumed in total. API cost for the single discovery: roughly $100,000.
The real bottleneck is now "human confirmation"
Two Anthropic researchers spent hundreds of hours brushing up on cryptography to validate the AES result. Claude produced the attack in a week; humans needed nearly a month to believe it.
That sentence weighs more than any benchmark score.
AI for Science is often framed as "models helping scientists," with the default assumption that scientists always retain judgment. But when models run beyond the boundary of human knowledge, researchers must simultaneously verify correctness, novelty, and real-world impact. As generation speed keeps climbing, the verification queue will overflow first.
Anthropic has partnered with academia on CryptanalysisBench and followed responsible disclosure, coordinating in advance with the HAWK authors, the relevant NIST community, and government and industry partners. The direction is right, but far from sufficient. If future models find practical vulnerabilities in production cryptographic systems, the cadence of disclosure, reproduction, and patching must be agreed in advance—not improvised on the spot.
My conclusion is simple: this didn't "break the internet," but it proves frontier models can already find new things in mathematical objects that experts have scrutinized for years. The research harness has matured. The next thing to bottleneck the industry will be the verification harness.
Original sources
- Anthropic research post: https://www.anthropic.com/research/discovering-cryptographic-weaknesses
- Technical demo repository: https://github.com/anthropics/cryptography-research-demo
- Related coverage: https://the-decoder.com/anthropic-says-its-mythos-model-found-vulnerabilities-in-cryptographic-algorithms-that-secure-the-internet