Anthropic’s AI model Mythos has uncovered a fundamental weakness in HAWK, a candidate algorithm for post-quantum cryptography. The discovery prompted the developers to withdraw HAWK from the standardization process led by the U.S. National Institute of Standards and Technology (NIST). While the attack does not pose an immediate risk to existing encryption, it does demonstrate that AI can play an increasingly significant role in cryptographic research, reports Ars Technica.
HAWK was one of the algorithms being evaluated by the U.S. National Institute of Standards and Technology (NIST) for future post-quantum standards. The digital signature algorithm had already passed two rounds of evaluation and was in a third evaluation phase, which specifically seeks out subtle vulnerabilities.
That vulnerability was discovered using Anthropic’s AI security model, Mythos. According to Ars Technica, the model was able to refine an existing attack on HAWK, reducing the effective key strength by roughly half. As a result, the algorithm no longer met the desired security margins, prompting the developers to withdraw it from the standardization process.
AI combines existing techniques
It is noteworthy that Mythos did not develop a new mathematical method. Instead, the model combined existing cryptographic techniques in a way that researchers had not previously applied.
Matthew Green, a professor of cryptography at Johns Hopkins University, considers this to be particularly remarkable. According to him, the AI did not introduce a new theory but rather brought together familiar tools in an unexpected way, with effective results.
The attack targets the so-called Lattice Isomorphism Problem, the mathematical foundation of HAWK. The vulnerability can be mitigated by doubling the key length, but this also significantly increases the computational power required. That makes HAWK less attractive than alternative post-quantum algorithms such as ML-DSA and FN-DSA.
Mythos was also used against AES, the world’s most widely used symmetric encryption standard. There, the model improved upon an existing man-in-the-middle attack using the Möbius Bridge technique.
According to Anthropic, this optimization can reduce the computation time required for this attack by a factor of 200 to 800. However, its practical significance remains limited. The attack was carried out exclusively on a weakened research variant of AES with seven encryption rounds. Production versions of AES use 10, 12, or 14 rounds and remain out of reach of this technique.
No threat to existing cryptography
Anthropic emphasizes that the results do not mean that current cryptographic standards have been cracked. Both HAWK and AES were examined using specially prepared test variants intended for scientific research. Furthermore, the attacks require substantial computing power and cannot be carried out outside a laboratory environment.
However, Anthropic does see this as an indication that large language models are becoming increasingly proficient at cryptanalysis. According to the company, this could lead to researchers spending more time in the future verifying vulnerabilities found by AI than discovering them.
At the same time, the question remains as to how much benefit Mythos has actually provided. HAWK was still a relatively new algorithm, and it cannot be ruled out that traditional cryptographers would have discovered the same weakness in the near term. For longer-established cryptosystems, such as RSA and elliptic-curve cryptography, Anthropic has not yet published comparable results.