NEW YORK — An advanced artificial intelligence model developed by Anthropic has identified previously unknown weaknesses in a simplified version of one of the internet’s most widely used encryption standards, raising new questions about AI’s long-term implications for cybersecurity, according to a new report.
The findings do not pose an immediate threat to online banking, financial transactions or encrypted communications because the AI cracked only a deliberately weakened version of the Advanced Encryption Standard (AES), not the version currently used to secure internet traffic, wireless networks and data storage, the New York Times reported.
However, researchers say the results demonstrate how rapidly AI is advancing in areas such as mathematics, coding and cybersecurity and suggest future AI systems could eventually challenge encryption methods long considered secure.
New Method of Attack
According to The New York Times, Anthropic’s experimental AI model, Claude Mythos Preview, discovered a new method for attacking the weakened encryption algorithm that made the process approximately 200 to 1,000 times faster than techniques previously developed by human researchers.
Researchers commonly study simplified versions of encryption algorithms to determine whether vulnerabilities could eventually extend to the stronger standards used in real-world systems.
Nicholas Carlini, an Anthropic research scientist who worked on the project, told The New York Times that AI capabilities have advanced dramatically in a short period.
“They could not do problems that I could do when I was 16,” Carlini said of AI models he tested last year. “Now they are doing state-of-the-art research not previously discovered in the field.”
Attack on HAWK
The newspaper also reported that Anthropic also used the AI system to develop an improved attack against HAWK, an experimental cryptographic system being considered by the National Institute of Standards and Technology as a possible future encryption standard designed to withstand attacks from both conventional and quantum computers.
According to Anthropic, the HAWK findings were validated by the system’s developers, while independent cryptographers reviewed the research involving the weakened AES algorithm before publication. The company said it shared its findings with the U.S. government and industry partners before releasing the research.
AI Initially Declined
The New York Times reported that the AI initially declined to analyze the encryption problem because it believed improvements were impossible. After researchers encouraged it to continue, the system spent about a week independently developing the new attack method, followed by nearly a month of human verification.
The research comes as governments increasingly scrutinize frontier AI systems because of their cybersecurity capabilities.
According to The New York Times, Anthropic restricted public access to Claude Mythos Preview after its April debut because of its ability to identify and exploit software vulnerabilities. The company has made the model available only to selected government agencies and organizations while federal officials evaluate its capabilities.
Escape from Controlled Environment
The newspaper also noted that OpenAI disclosed last week that one of its AI models escaped a controlled testing environment and accessed an internal digital library to obtain information that would improve its performance on evaluation tests, adding to concerns about increasingly autonomous AI behavior.
Cybersecurity experts have long warned that if widely used encryption standards were ever compromised, the consequences for digital privacy, financial services and national security could be severe. Intelligence agencies have also expressed concern that nations such as China may be collecting encrypted communications today in hopes that future advances in computing or AI could eventually decrypt them.
Closer Attention Urged
Although experts emphasized there is no immediate threat to today’s encryption systems, some believe the pace of AI development warrants close attention.
“Given that we are constantly underestimating the power and time of availability of future models, are we really comfortable that two years from now strong encryption won’t be threatened?” Glenn S. Gerstell, former general counsel of the National Security Agency, told The New York Times.
Gerstell said mathematicians generally believe today’s encryption remains secure, but cautioned against dismissing the potential capabilities of future AI systems before quantum-resistant cryptography becomes widespread.As the CU Daily reported here, during the recent World Credit Union Conference, America’s Credit Unions’ President and CEO Scott Simpson also expressed concerns over the risks from quantum computing, stating, “I do not believe any government on Earth is devoting enough policy attention or financial resources to the issue. I think
