Anthropic AI model identifies security exploits in major software platforms
A research model from Anthropic has successfully generated multi-step exploit chains targeting Chrome, Android, and Windows. All identified vulnerabilities have been patched by the respective vendors.
Anthropic’s Frontier Red Team reported that an internal research model successfully generated functional exploit chains across widely used software platforms. The model was able to identify and connect multiple individual vulnerabilities—ranging from sandbox escapes in Google Chrome to data extraction methods on locked Android devices—to form complete attack sequences.
In the case of Chrome, the AI identified four distinct vulnerabilities that enabled arbitrary code execution outside the browser's security sandbox. Other findings included a method to bypass USB data restrictions on locked Android devices and a technique to crash Windows servers remotely via the HTTP/3 protocol. The research involved the model performing reverse engineering on binaries and analyzing open-source codebases.
Anthropic stated that it disclosed all findings to the respective vendors before releasing any information. All vulnerabilities mentioned in the report have since been resolved through patches issued by Google, Apple, and Microsoft. The findings highlight the evolving capability of AI systems to assist in complex security research and automated vulnerability discovery.