Uncovering Cryptographic Vulnerabilities via Claude

Executive Summary: Researchers at Anthropic, utilizing the Claude Mythos Preview model, have identified advanced methods for attacking cryptographic algorithms. These include a significant weakening of the post-quantum signature scheme HAWK and a novel attack on round-reduced AES. While these represent major research milestones, they do not currently compromise any live production environments.

---

Introduction: A Shift in AI Capability

When Claude Mythos Preview was first introduced, its primary achievement was the autonomous discovery and exploitation of software vulnerabilities. Previously, Claude's successes in cryptographic libraries were limited to implementation errors—essentially, mistakes made by programmers when coding an algorithm.

However, the latest research reveals a critical evolution: Claude can now identify fundamental mathematical flaws within the algorithms themselves.

Understanding the Stakes

Cryptography is the bedrock of digital trust. To visualize the layers of security used when visiting a site like https://www.anthropic.com, consider this flow:

If these mathematical foundations fail, billions of users' private communications, banking details, and emails become vulnerable to interception.

---

Key Research Findings

1. Weakening the HAWK Signature Scheme

The US National Institute of Standards and Technology (NIST) is currently seeking "post-quantum" algorithms—systems that can withstand the processing power of future quantum computers. HAWK is a third-round candidate in this selection process.

Despite two years of rigorous human expert review, Mythos improved the best-known attack on HAWK in only 60 hours. This discovery effectively reduced the security margin:

$\text{Effective Key Strength} \approx \frac{1}{2} \times \text{Original Strength}$

2. Attacking Round-Reduced AES

The Advanced Encryption Standard (AES) is the global gold standard for symmetric encryption. To test its limits, researchers often study "round-reduced" versions (simplified versions of the cipher).

Mythos discovered a way to break one of these weaker versions by eliminating a necessary attacker guess. This resulted in a massive performance boost for the attack:

• Speed Increase: $200\times$ to $800\times$ faster than previous best-known methods.

---

Impact Assessment & Safety

It is important to emphasize that no current production systems are at risk.

• HAWK: Not yet deployed (still a candidate).
• AES: The attack only applies to a reduced version, not the full cipher.
• Production Software: No changes required.

This process is an example of cryptanalysis working as intended: by stress-testing algorithms, we build more resilient systems.

---

Methodology and Costs

The discovery process was largely autonomous. While one researcher collaborated with Claude on the HAWK attack, the AES attack was achieved via a "scaffold"—a programmatic framework that allowed the AI to operate independently.

Conceptual Scaffold Logic:

while vulnerability_not_found:
    hypothesis = claude.generate_attack_vector(cipher_specs)
    result = test_environment.execute(hypothesis)
    claude.refine_strategy(result)

Resource Investment:

• Human Effort: Minimal (mostly autonomous).
• Financial Cost: Approximately $100,000 in API costs per discovery.

---

Collaboration and Responsible Disclosure

To ensure these findings benefit the wider community, Anthropic took several steps:

1. Academic Partnership: Collaborated with the University of Haifa, Tel Aviv University, and ETH Zurich to create CryptanalysisBench, a standardized benchmark for evaluating LLM cryptanalytic skills.
2. Government Liaison: Shared findings with US industry and government partners.
3. Direct Disclosure:
   • Notified HAWK authors in June.
   • Coordinated public disclosure via the NIST mailing list.

Detailed technical breakdowns of these attacks are available in two newly published papers, with further results expected soon.