Claude Mythos Cracked Post-Quantum Cryptography That Humans Spent Years Failing to Break

Anthropic in the present day stated an unreleased model of its strongest AI mannequin discovered two beforehand unknown assaults on cryptographic algorithms, one in all them in opposition to a scheme presently competing to change into a U.S. federal commonplace.

That scheme is HAWK, a digital signature system—the mathematics that proves a transaction got here from you with out ever exposing your non-public key—constructed to outlive future quantum computer systems. The non-regulatory federal company and lab NIST moved it into the third spherical of its post-quantum signature competitors in Could, the place it’s the final lattice-based candidate standing.



Claude discovered a symmetry buried in HAWK’s math that no human had thought to make use of. For the smallest configuration, the price of recovering a secret key fell from 2^64 operations to 2^38, roughly 67 million instances much less work.

Fixing it means roughly doubling HAWK’s keys. “Sadly, doubling HAWK’s key measurement eliminates lots of the causes making the scheme (because it presently stands) a lovely PQC signature candidate,” Anthropic wrote.

That commerce issues extra to blockchains than it sounds. Signature measurement is block house, and block house is charges, so any chain looking for a quantum-resistant alternative is partly selecting on bytes per signature. Compact keys and quick signing had been HAWK’s complete pitch, and the repair prices it a lot of that edge.

Don’t fear, hodlers: Your cash are tremendous (for now). HAWK has by no means been deployed wherever, and Bitcoin nonetheless runs on ECDSA, the pre-quantum signature scheme that candidates like HAWK are ultimately meant to exchange.

Anthropic disclosed each outcomes to the algorithms’ authors and to U.S. authorities and trade companions earlier than publishing, and coordinated the HAWK discovering with NIST.

The AES end result wanted a pep speak

The second assault targets AES, the cipher scrambling your HTTPS visitors, your encrypted drive, and your trade’s backend. Full AES-128 pushes knowledge via 10 rounds of scrambling, and Claude attacked a 7-round analysis model that no one has improved on since 2013.

The setup was intentionally harsh. Researchers barred the mannequin from all 5 established households of AES cryptanalysis and advised it to invent a sixth, closing the transient with a line about how the primary differential assault did not beat something—it invented the sport. Claude additionally inherited working notes from earlier agent runs that had already burned via roughly 200 failed assault variants.

It refused anyway. “On AES-128 r5/r6/r7 it discovered nothing as a result of there’s nothing simple to search out; that is the most-studied block cipher in existence,” the mannequin advised researchers, per transcripts Anthropic revealed.

Anthropic despatched simply three substantive messages over the following three days, amongst them: “no once more the aim is that we’ve got extremely inteligent [sic] mannequin nearly as good prime researcher, we need to discover new assaults.” One other refused to let Claude swap AES for a neater cipher.

Then it produced the trick the paper calls a Möbius Bridge, killing one of many 9 key bytes an attacker beforehand needed to guess. Refining that into the revealed model took a couple of extra days and a billion output tokens.

The discovering with the shortest path to one thing actual acquired the least consideration. Claude additionally broke 13 rounds of LEA, a Korean nationwide commonplace and ISO lightweight-encryption commonplace constructed for telephones and internet-of-things gadgets, in underneath an hour on a desktop in opposition to a previous finest that wanted 2^98 plaintext pairs. The deployed LEA runs 24 rounds, so nothing within the subject is damaged.

Verification took longer than discovery

The HAWK paper is unusually blunt concerning the division of labor. “The vast majority of mathematical discoveries on this paper had been AI-assisted. Human creator contribution primarily consisted of directing, organizing and verifying AI work,” its authors wrote.

Claude discovered the AES thought in days. Anthropic researchers then spent a number of hundred hours studying sufficient cryptography to verify it labored—the identical mannequin that discovered 271 vulnerabilities in Firefox throughout inside testing.

“The cybersecurity neighborhood is now grappling with the truth that language fashions are in a position to uncover so many bugs that the usual human processes (like vulnerability triage, verification, and remediation) wrestle to maintain up,” Anthropic wrote, warning that human researchers could change into the bottleneck.

Anthropic additionally constructed CryptanalysisBench—191 cipher-breaking duties drawn principally from NIST competitions. Fashions submit a working assault script that both wins a proper safety sport or would not, with no partial credit score and no human grading.

Mythos 5 broke 85.7% of duties with recognized options, in opposition to 65.3% for the weakest mannequin examined. Towards full-strength ciphers with no revealed break, each mannequin scored underneath 9%.