Cryptography breakthrough: claude cracks hawk and weak AES variant

The Frontier Red Team at Anthropic recently published findings on two cryptographic algorithms: HAWK, a post-quantum signature scheme, and a reduced version of AES. While these discoveries do not pose immediate threats to current systems, they highlight vulnerabilities that warrant attention. Here’s what you need to know.

Post-quantum falcon falters

HAWK was a strong contender in the NIST’s post-quantum cryptography standardization process, designed to resist quantum computing threats. After passing two rounds of human expert evaluation over two years, Mythos—a cutting-edge AI model—identified a critical weakness.

The breakthrough hinged on discovering a geometric symmetry in HAWK’s underlying module Lattice Isomorphism Problem. This allowed Mythos to reduce the key recovery complexity from 2⁶⁴ to just 2³⁸ operations. In a real-world test, the algorithm’s secret key was extracted in just 3 hours and 42 minutes using a standard 96-core server.

The implications are clear: HAWK’s security margin is now compromised. To restore acceptable protection levels, the algorithm’s key and signature sizes would need to double, negating its primary advantage of computational efficiency. Consequently, HAWK’s developers have withdrawn it from the NIST selection process, marking a significant setback for post-quantum cryptography.

Why this matters

• Quantum resistance: HAWK’s failure underscores the fragility of post-quantum algorithms still in development.
• Process validation: Even algorithms vetted by experts can harbor hidden flaws, as Mythos demonstrated.
• Future-proofing: The need to double key sizes for acceptable security adds computational overhead, complicating adoption.

What about AES? A theoretical but notable crack

AES, the gold standard for symmetric encryption, remains unscathed in its full 128-bit or 256-bit forms. However, Mythos uncovered a vulnerability in a deliberately weakened variant of AES-128, reduced to just 7 rounds instead of the standard 10.

The attack leveraged a novel Möbius Bridge technique—an invariant fingerprinting method integrated into a meet-in-the-middle approach. This innovation slashed the computational effort of prior attacks by a factor of 200 to 800, though it remains confined to academic exploration.

Anthropic’s researchers emphasized that the attack’s practical implications are negligible. The reduced AES variant is not used in operational systems, and the full AES standard’s additional rounds render it immune to such exploits. The French National Agency for the Security of Information Systems (ANSSI) continues to recommend AES-128 as a baseline, with AES-192 or AES-256 for post-quantum scenarios.

Key takeaways

• No immediate risk: AES in its standard form is unaffected.
• Theoretical advances: The attack on the reduced AES variant showcases AI’s potential to uncover subtle cryptographic weaknesses.
• Preparation for the future: Organizations should consider transitioning to larger key sizes (e.g., AES-192 or AES-256) as a precaution against emerging threats.

Reflections on AI’s role in cryptanalysis

Anthropic’s findings underscore a paradigm shift in cryptography research. While HAWK’s crack required approximately 60 hours of AI-driven analysis—costing an estimated $100,000 in API fees—it demonstrates how AI can accelerate vulnerability discovery. Mythos initially resisted the task, arguing that AES’s robustness made improvements impossible. After fine-tuning the model’s parameters (a process Anthropic calls “scaffolding”), the AI generated a billion tokens over several days to devise the new attack method.

This episode also highlights a growing challenge: the cybersecurity community now faces a bottleneck not in finding vulnerabilities but in validating AI-generated discoveries. Anthropic’s team spent a month rigorously verifying Mythos’ results, ensuring mathematical rigor and eliminating potential hallucinations. This human-AI collaboration is critical to maintaining trust in cryptographic research.

What’s next?

The cryptography landscape is evolving rapidly. While HAWK’s withdrawal and AES’s theoretical vulnerability do not disrupt current systems, they serve as a reminder of the importance of continuous scrutiny. Anthropic has hinted at further discoveries, framing them as “preliminary experiments.” The message is clear: more breakthroughs are on the horizon, and the cybersecurity community must adapt to this new era of AI-assisted cryptanalysis.

For now, the message remains reassuring: there’s no cause for panic. But the conversation around cryptographic resilience is far from over.