Anthropic has announced that its unreleased Claude Mythos Preview artificial intelligence model identified previously unknown mathematical weaknesses in reduced-round versions of the Advanced Encryption Standard (AES), one of the world’s most widely used encryption systems. While the breakthrough does not compromise the full AES encryption protecting banking, communications, Wi-Fi, and government systems today, security experts view the achievement as a significant demonstration of AI’s rapidly expanding capability in advanced cryptanalysis. The research also uncovered weaknesses in the HAWK post-quantum signature scheme, adding urgency to ongoing efforts to strengthen next-generation cryptographic standards. The findings underscore that AI is increasingly capable of performing sophisticated mathematical research once thought to require years of work by elite human cryptographers, forcing governments, technology companies, and cybersecurity professionals to accelerate preparations for an era in which AI itself becomes both a powerful defensive tool and a potentially formidable offensive capability.
Sources
- https://www.nytimes.com/2026/07/28/us/politics/anthropic-ai-encryption-security-aes.html
- https://www.anthropic.com/research/discovering-cryptographic-weaknesses-with-claude
- https://arstechnica.com/security/2026/07/mythos-attack-on-3rd-round-pqc-algorithm-candidate-puts-it-out-of-commission
Key Takeaways
- AI has now demonstrated the ability to discover meaningful cryptographic weaknesses that traditionally required years of specialized human research, signaling a major shift in cybersecurity research capabilities.
- Current internet encryption has not been broken; the discoveries involve reduced-round AES and a post-quantum candidate algorithm rather than the full AES implementations protecting everyday online transactions.
- The development increases pressure on governments and industry to modernize cryptographic standards, expand AI-assisted defensive research, and prepare for a future in which advanced AI dramatically accelerates both cyber offense and cyber defense.
In-Depth
For decades, modern encryption has been regarded as one of the strongest foundations of the digital economy, protecting financial transactions, government communications, military systems, and the private information of billions of people. The latest research demonstrates that artificial intelligence is beginning to challenge long-held assumptions about how quickly sophisticated cryptographic analysis can be performed. Although the findings do not mean online banking or encrypted messaging has suddenly become vulnerable, they do illustrate how rapidly frontier AI systems are advancing into areas once reserved for small communities of highly specialized mathematicians.
The broader implication extends well beyond a single encryption algorithm. As AI systems become capable of conducting original mathematical research, the pace at which both vulnerabilities and defenses emerge is likely to accelerate dramatically. That reality presents policymakers with a familiar challenge: technological innovation is advancing much faster than regulatory and security institutions typically adapt. Rather than reacting after vulnerabilities become practical threats, governments and private industry will need to invest aggressively in stronger cryptographic standards, AI-assisted defensive tools, and continuous security testing.
The episode also serves as a reminder that technological leadership carries national security consequences. Nations capable of developing frontier AI models will likely enjoy growing advantages in cybersecurity, intelligence, and critical infrastructure protection. Ensuring that those capabilities remain aligned with democratic institutions, responsible disclosure practices, and robust security standards may prove just as important as the underlying breakthroughs themselves.