Thought Leadership
An unreleased research model from Anthropic has uncovered serious security vulnerabilities in Chrome, Safari, Android, Windows, and Linux, according to the company.
The findings show that modern AI models can combine multiple vulnerabilities into attack chains that can enable code execution and the theft of sensitive data, among other threats.
Chaining vulnerabilities into working exploits
In a study, Anthropic describes five attack scenarios in which its internal Claude model not only identified vulnerabilities but also combined them to build working exploits. In several cases, the attacks required little or no additional user interaction.
In Chrome, the model successfully chained multiple vulnerabilities to execute code outside the browser’s security sandbox. As a proof of concept, an exploit running on a Mac with a standard Chrome configuration launched the Calculator app after a webpage was loaded.
Safari was also affected. There, the model developed an attack chain that bypassed the Same-Origin Policy, a security mechanism designed to prevent websites from accessing data belonging to other websites. A successful attack could potentially expose browsing history, cookies, or authenticated sessions.
Risks for mobile devices and servers
On Android, Claude reportedly identified a way to extract photos and other personal information from a locked smartphone over a USB connection. Another vulnerability involved the handling of HTTP/3 requests on Windows. It could allow attackers to repeatedly crash vulnerable servers remotely.
Anthropic considers its Linux finding particularly serious. According to the company, the model developed an exploit that could allow a remote attacker to execute commands with root privileges on a server running a Multipath TCP listener. The attack would require neither credentials nor user interaction. These examples demonstrate that combining multiple flaws can have significantly more severe consequences than examining individual vulnerabilities in isolation.
AI lowers the barrier to exploit development
Anthropic says the findings suggest that advanced AI models could make the development of sophisticated attacks accessible to a much broader range of people. “Frontier models have the ability to discover and combine multiple vulnerabilities into a working exploit,” the company writes. Until now, the ability to develop such attack chains has largely been the preserve of specialized experts.
According to Anthropic, the discovered vulnerabilities were reported to the affected vendors through a coordinated disclosure process. The company also points to its Cyber Verification Program, which aims to help identify and fix security flaws before attackers can exploit them.
OpenAI also reports hundreds of vulnerabilities
Anthropic is not the only company to have used unreleased AI models to uncover security flaws. The previous month, OpenAI reported that an internal cybersecurity model had identified hundreds of vulnerabilities. These reportedly included flaws in Chrome’s V8 JavaScript engine, a mobile operating system, a database, and an operating system kernel.
Both companies see an opportunity to accelerate security research and uncover flaws before cybercriminals can exploit them. However, identifying a vulnerability is only the first step. Security experts must then validate the findings, vendors need to develop appropriate fixes, and administrators must install the patches on affected systems.
The window for mitigation could be shrinking
As AI models become more capable, the time between discovering a vulnerability and its potential exploitation could shrink. This puts additional pressure on organizations to assess security disclosures quickly and deploy updates without delay. The challenge is therefore not just finding vulnerabilities. Vendors and system operators must also keep pace with the speed of automated analysis. AI can strengthen security research, but it can also accelerate the development of sophisticated attack techniques.
(Editorial Team)
