AI company Anthropic has declared that its unannounced model, Claude Mythos Preview, has discovered formerly unknown mathematical weaknesses in cryptographic algorithms that human researchers had missed for years.
Cryptography uses sophisticated algorithms to protect web traffic, e-mail, software updates and sensitive information like banking details. It does this by rendering the data entirely unreadable.
The Anthropic team explains its work in two papers.
In the first paper, the team tested a prototype security approach referred to as HAWK, that could emerge as a future global standard for the era of ultra-powerful quantum computers.
Human experts had spent more than 2-years analyzing the algorithm and believed it was secure. Moreover, the AI model managed to discover a hidden mathematical shortcut that made it much simpler to crack than researchers had earlier believed. It found out a new way to attack the system that researchers had overlooked.
Although AI did the heavy mechanical lifting to discover this flaw, humans acted as supervisors, as Anthropic researchers Zygimantas Straznickas and Stephen A. Weis point out of their paper, “The majority of mathematical discoveries in this paper were AI-assisted. Human author contribution in particular consisted of directing, organizing and verifying AI work.”
In the second paper, researchers Milad Nasr and Nicholas Carlini examined limited versions of AES (Encryption Standard), which is the standard system for defensive internet data and communication security today. Cryptographers often study those scaled-down versions to see if the overall system has any hidden weaknesses. Moreover, the best-known method for analyzing one of these reduced versions had stayed unchanged since 2013.
Faster AES testing
Working entirely on its own, the AI model invented a mathematical shortcut called a Möbius Bridge. This removes a trial-and-error segment that human researchers usually perform, rushing up the whole testing procedure by around 200 to 800 times compared with preceding human methods.
“These upgrades have been all observed entirely autonomously by a large language model,” the researchers commented in their paper. What’s more, this new system is simple to verify. “The Möbius transform is quite straightforward to audit, and we’re extremely confident it’s correct.”
Protecting tomorrow’s security landscape
The two papers from Anthropic do not mean that recent encryption standards are vulnerable or that sensitive personal or government data are all at once unsafe. Moreover, they display that AI can assist researchers find out previously unknown mathematical weaknesses and improve the analysis of cryptographic systems earlier than those flaws become real-world security issues. This could permit us to build even stronger defenses towards future cyberattacks.
