Sixty hours to find the attack, one month to believe it

Anthropic's Claude Mythos model finds security vulnerabilities in a matter of hours. Human validation time is becoming the bottleneck of research.

Two cryptanalysis works have emerged from Anthropic's Frontier Red Team, conducted using Claude Mythos Preview. The first targets HAWK, a digital signature scheme candidate in the third round of the NIST post-quantum cryptography call. After two years of peer review by human experts, the model improved the best known attack in about sixty hours, by exploiting a previously unused lattice symmetry. The key sizes would need to be doubled to restore the claimed security level, which strips the scheme of much of what made it an attractive candidate. The result is specific to HAWK and does not affect other NIST post-quantum proposals.

The second focuses on a weakened version of AES, reduced to seven rounds out of ten. A signature technique dubbed Möbius Bridge eliminates one of the enumerations the attacker had to perform, making the existing attack 200 to 800 times faster. Neither of the two works affects production systems: HAWK is not deployed anywhere, and the targeted AES variant assumes a volume of queries far beyond any practical reach. Anthropic itself describes these two results as expected.

The process is worth noting. On AES, the model initially refused to engage, taking for granted that nothing could be improved on the most heavily studied block cipher in existence. It had to be prompted repeatedly, through messages that the company publishes as-is, typos included, for it to rewrite its own harness and start looking for genuinely new ideas. Each result cost approximately one hundred thousand dollars in API calls.

The most telling point lies downstream. While the model found the attack on AES in one week, two researchers, who were not cryptography specialists, took nearly a month to convince themselves that it was valid. The company anticipates that human validation will become the limiting factor in cryptographic research, as it has already become for software vulnerability triage. Both flaws were disclosed to the respective authors, the public NIST mailing list, and US government and industry partners prior to publication. A benchmark, CryptanalysisBench, is being released in parallel with academics from ETH Zurich, Tel Aviv University, and the University of Haifa.