claude cryptography ai-safety red-teaming

Discovering Cryptographic Weaknesses with Claude

anthropic.com

July 28, 2026

18 min read

🔥🔥🔥🔥🔥

59/100

Summary

Researchers at Anthropic have utilized Claude Mythos Preview to uncover vulnerabilities in cryptographic algorithms. One attack significantly compromises the HAWK digital signature scheme, while another reveals a new method to exploit round-reduced AES, a widely used symmetric cipher.

Key Takeaways

Researchers at Anthropic used Claude Mythos Preview to discover improved attacks on cryptographic algorithms, specifically weakening the HAWK digital signature scheme and identifying a new attack on round-reduced AES.
The attack on HAWK effectively cut its key strength in half within 60 hours of work, despite it being a candidate for post-quantum security.
Claude Mythos Preview improved the speed of attacks on a weaker version of AES by 200-800 times, showcasing the potential of AI in identifying flaws in cryptographic systems.
These findings do not currently impact production systems, as HAWK is not deployed and the attack on AES concerns a reduced version of the algorithm.

Read original article

View story →

Coverage & discussions

Hacker News · anthropic.com

Discovering Cryptographic Weaknesses with Claude ↗

Community discussions

Hacker News discussion ↗

Historical count: 5 · 5 saved in sample · Historical sample

Analysis ready

Community Sentiment

Mixed

Positives

Spending $100k in tokens in a week is an impressive feat even with massive parallelization, showcasing the power of Claude's capabilities in cryptographic research.
The collaborative approach led to significant breakthroughs in discovering vulnerabilities, highlighting the potential of multi-agent workflows in AI applications.
The HAWK attack's ability to halve the effective key length is a noteworthy achievement in cryptographic research, raising questions about future standards.

Concerns

The obsession with 'prompt engineering' feels misguided when the software still struggles to follow instructions effectively, pointing to fundamental limitations in LLMs.
There's a growing tech aristocracy with access to powerful AI tools that could widen the gap between those who can afford extensive API usage and the average user.