Artificial intelligence can create software, check IT systems or clone itself. Now a new AI model from the US company Anthropic is causing a stir. Because “Claude Mythos Preview” discovered thousands of previously unrecognized vulnerabilities in core components of the global IT infrastructure. Components of the Linux operating systems used by servers worldwide, as well as web browsers and software for firewalls, routers and VPN systems, are affected. Anthropic thinks this is so dicey that they are not publishing their AI model for the time being. Instead, they have selectively granted only major IT infrastructure operators access to their AI so that they can fix the vulnerabilities in their codes. But how big a threat does advanced AI pose to cybersecurity?
Artificial intelligence is developing rapidly. The most advanced AI systems, so-called frontier models, already achieve performance in some areas that is equal to or even superior to that of most people – this also applies to the IT sector. AI models like Claude Code require little help from human programmers to create software and IT applications using just a few targeted prompts. The use of artificial intelligence in the IT sector is already widespread. “Frontier models already play a role in cyber attacks and their defense: They can accelerate attacks – for example when analyzing vulnerabilities and their exploitation. At the same time, they can strengthen defenses – also when analyzing code for possible vulnerabilities. They also help with closing possible vulnerabilities or analyzing attacks,” explains Thorsten Holz from the Max Planck Institute for Security and Privacy in Bochum.
Zero-day vulnerabilities in Linux, firewalls and co
A new AI model from the US company Anthropic now shows what this means in concrete terms. With “Claude Mythos Preview” it has developed an artificial intelligence that, in its own assessment, can become a threat to global cybersecurity. “Claude Mythos Preview reveals a threatening fact: AI models have reached a level of coding skill that surpasses almost all humans in finding and exploiting software vulnerabilities,” Anthropic explains in a blog post. This assessment is based on internal tests in which the AI system discovered serious security gaps in globally distributed software. “Over the past few weeks, we have used Claude Mythos Preview to identify thousands of zero-day vulnerabilities – including many critical ones – in every major operating system and web browser, and other major systems,” Anthropic reports. Zero day exploits are security gaps in software that were previously undetected and that can be exploited by hackers.
It is nothing new that AI models can find vulnerabilities in existing computer code. However, there is concern about the efficiency and speed with which the new AI model finds such potential attack targets: “Claude Mythos Preview demonstrates a leap in these cyber skills,” explains Anthropic. “The vulnerabilities that AI discovered have, in some cases, survived millions of automated security tests and decades of human review.” As an example of such security gaps, Anthropic cites a vulnerability in the OpenBSD open source system that has already been fixed and remained undetected for 27 years, which is found in firewalls, routers, web browsers and VPN programs, among other things. “The vulnerability allowed attackers to remotely crash any machine,” said Anthropic. Another example is a chain of multiple vulnerabilities in the Linux kernel, the basis of almost all server software. In the worst case, a hacker could gain control of server systems through these attack points. This security gap has now also been closed.
(Video: Anthropic)
New scale of cyber attacks
In Anthropic’s opinion, these examples and the thousands of security vulnerabilities that are still kept secret demonstrate the threat that advanced AI models like Claude Mythos Preview can pose to the world’s IT infrastructure. Because if such AI systems become freely accessible, hackers could use them to cause serious damage, as Jörn Müller-Quade, Professor of Cryptography and IT Security at the Karlsruhe Institute of Technology (KIT), explains: “Attack options such as those found by Anthropic’s new model are nothing new. What is new, however, is how cyberattacks are now scaling: previously it required very skilled hackers to find such vulnerabilities. Now an AI model allows any layperson to carry out attacks In the future, we can no longer afford to develop software that accepts errors and only corrects them gradually. This applies at least to critical infrastructures.”
In order to prevent the threat, at least in part, Anthropic has decided not to publish Claude Mythos Preview for the time being. Instead, they initiated the “Glasswing” project: 40 large software companies receive access to the AI model so that they can fix potential vulnerabilities in their systems. Those involved include Microsoft, Google, the Linux Foundation, NVIDIA, Apple, Amazon Web Services, Cisco, CrowdStrike, the Apache Foundation, Broadcom and JPMorganChase. Open source organizations such as Linux and Apache receive additional millions of dollars in donations to support this effort. At the same time, the company is developing new security measures for the next version of the Claude Opus AI model that are intended to limit its IT capabilities and thus prevent misuse.
“The Glasswing project is just a beginning,” emphasizes Anthropic. The cybersecurity problems caused by artificial intelligence cannot be solved by one organization or company alone. “AI developers, other software companies and open source operators, IT security researchers and governments all have a role to play. We must act now to ensure cybersecurity reigns supreme.”
Source: Anthropic, Science Media Center