On July 28, 2026, Anthropic’s Frontier Red Team published a research note titled “discovering cryptographic weaknesses with Claude”. The findings, derived from an internal preview of Mythos not yet available to the public, focus on two security algorithms: HAWK, a post-quantum digital signature scheme, and a reduced version of AES, commonly used in academic research to test cryptographic robustness.
Anthropic emphasizes that neither discovery poses a threat to current computing systems. However, the implications—while far from catastrophic—are not negligible, contrary to what some headlines might suggest.
Post-quantum falcon: HAWK’s brief flight
HAWK was a strong contender in the NIST’s post-quantum cryptography standardization process, designed to complement the four algorithms selected in 2022. It had successfully passed two rounds of expert review over two years—until Mythos uncovered its vulnerability.
This isn’t the first time a promising post-quantum candidate has fallen short. SIKE, for instance, can be broken in under an hour on a standard PC, while Rainbow, another finalist, had its secret key exposed after an average of 53 hours. HAWK’s downfall follows a similar pattern.
The attack leveraged Mythos’s ability to exploit geometric symmetries in the lattice isomorphism problem underpinning HAWK’s security. By identifying an automorphism, researchers reduced the private key recovery to a short vector problem in a lattice with halved dimensions. Concretely, for HAWK-256, the complexity dropped from 2⁶⁴ to 2³⁸ operations, enabling key extraction in just 3 hours and 42 minutes on a standard 96-core server.
While the attack doesn’t render HAWK unusable, it severely undermines its practicality. Restoring acceptable security levels would require doubling key and signature sizes, negating the algorithm’s main advantage: computational efficiency. As a result, HAWK’s designers officially withdrew it from NIST’s selection process on July 29, effectively rendering it obsolete.
It’s worth noting that symmetric algorithms like AES require doubling key sizes to resist quantum attacks—a coincidence that highlights the broader challenges in post-quantum cryptography.
What’s really happening with AES?
AES, the widely adopted encryption standard, powers everything from banking transactions to TLS protocols. Hardware acceleration via Intel and AMD instructions has made it a cornerstone of modern cryptography. Yet, even AES isn’t immune to scrutiny—particularly in academic settings, where researchers test weakened versions to probe its limits.
Mythos’s breakthrough targets a simplified AES-128 variant, reduced to seven rounds instead of the standard ten. The innovation lies in the “Möbius Bridge,” an invariant hashing technique integrated into a meet-in-the-middle attack. This method eliminates a previously unavoidable exhaustive search step (256 iterations), accelerating existing attacks by a factor of 200 to 800.
Anthropic is quick to clarify: this vulnerability is purely theoretical. The attack remains impractical against full-round AES, where three additional transformation rounds render it infeasible. The ANSSI’s latest cryptographic guidelines already recommend 192-bit keys for post-quantum security, with minimal performance overhead compared to AES-128.
Food for thought
While neither discovery affects real-world systems today, the implications extend beyond HAWK’s withdrawal or AES’s theoretical vulnerabilities. The episode underscores a growing challenge: the speed at which AI-driven tools like Mythos uncover flaws now outpaces human verification capabilities.
Anthropic’s research highlights the computational cost of such discoveries. Breaking HAWK required 60 hours of autonomous computation, with an estimated API cost of $100,000. Similarly, the AES attack demanded a billion tokens over several days, with the same price tag under standard conditions. Before publication, two researchers spent a month verifying the findings, eliminating potential hallucinations and ensuring mathematical rigor.
This raises critical questions: Are we entering an era where AI’s pace of discovery outstrips our ability to validate results? The cryptographic community is already stretched thin, with Firefox, Windows, and Apple issuing record-breaking patches. The bottleneck has shifted from finding flaws to verifying them—a challenge that grows exponentially with each breakthrough.
Anthropic’s work also reveals the iterative nature of AI-driven research. The Mythos preview initially refused to engage with the reduced AES variant, claiming no improvements were possible. After refining the prompts, the model autonomously redesigned its attack framework, demonstrating how AI can push the boundaries of cryptanalysis—even when conventional wisdom suggests a dead end.
The findings are part of a broader trend. Anthropic hints at “preliminary experiments” on other algorithms, signaling that more revelations are on the horizon. The message is clear: AI is reshaping cryptography, and the community must adapt to keep pace.
