Mythos Shows Feasible Attack on Early‑Stage Post‑Quantum Cipher, Undermining Its Viability

Mythos unveiled a novel meet‑in‑the‑middle technique that leverages a Möbius Bridge—a sophisticated fingerprinting algorithm. By employing this method, the code produced by Mythos cut the number of required inputs to 289, which Anthropic says can accelerate such attacks by a factor of 200 to 800.

Generating that many inputs pushes the attack far beyond what can be realistically executed outside a laboratory. Moreover, the speed‑up measured used a weakened version of AES with only seven rounds; standard AES implementations use ten, twelve, or fourteen rounds depending on key size, so the real‑world impact remains uncertain.

Anthropic’s blog post emphasizes these caveats but still argues that the results could fundamentally reshape cryptanalysis. The researchers note that language models are now discovering vulnerabilities at a scale that challenges traditional human‑driven processes such as triage, verification, and remediation.

“The cybersecurity community is now grappling with the fact that language models are able to discover so many bugs that the standard human processes (like vulnerability triage, verification, and remediation) struggle to keep up,” Anthropic wrote. “We predict that the same will soon be true in academic cryptography research. As language models increasingly produce novel research outputs autonomously, human researchers may become bottlenecked on studying and validating these results for technical validity, novelty, and utility.”

The report does not disclose whether Mythos was applied to more established cryptosystems such as elliptic‑curve cryptography or RSA. Achieving the most dramatic result against an algorithm still in its early stages leaves open the question of how much advantage Mythos truly offers compared with conventional cryptanalysis techniques.

Ultimately, the key takeaway is that AI‑assisted cryptanalysis remains unproven, and providers may be inclined to overstate its benefits. Nonetheless, there is growing evidence that large language models can provide significant advantages in uncovering cryptographic weaknesses, suggesting they may one day play a critical role in the race between securing and compromising vital assets.

The headline and body of this story have been updated to reflect the withdrawing of HAWK.

Source link