Agentic data science with the Data Agent Kit
cloud.google.com·7h ago
Data Science
Simon Willison's Blog·Jul 28·5 min read
Discovering cryptographic weaknesses with Claude
TL;DR
Anthropic researchers utilized Claude Mythos to identify mathematical flaws in the HAWK cryptographic scheme and a weaker version of AES, revealing that effective prompting is crucial for the model's success in this domain.
✦ Why It Matters
Review your cryptographic implementations and consider using AI tools for vulnerability assessments.
Key Takeaways
- Explore advanced AI models for cryptographic research.
- Encourage models to seek significant vulnerabilities.
- Invest in API usage for extensive research.
- Focus on meaningful contributions rather than easy wins.
Full Summary
Anthropic's Claude Mythos was employed to explore cryptographic weaknesses in HAWK and a less secure variant of AES (Advanced Encryption Standard). The research aimed to push the boundaries of cryptography by encouraging the model to seek out significant vulnerabilities rather than settling for easier, less impactful findings.
Over a span of 60 hours, costing approximately $100,000 in API usage, the team provided specific prompts to motivate the model, emphasizing the need for substantial contributions to the field. Despite the effort, the identified flaws do not pose immediate threats to existing computer systems.
This work highlights the potential of AI in cryptographic research while also illustrating the challenges in extracting meaningful results from advanced models.