Unreleased Anthropic AI creates functional exploits from security vulnerabilities
October 31, 2026 · 2 min read
An unreleased artificial intelligence from Anthropic identified security flaws in Windows, Linux, Chrome, and Safari and combined these defects to create functional invasion codes. The Frontier Red Team, which is the security team at the company, revealed these findings in a report published on Tuesday, October 27.
Researchers stated that the attack codes were constructed in a laboratory setting following initial commands given to the model. There are no records of real attacks exploiting these flaws on the internet. In one experiment, the technology opened the calculator on a Mac computer using the Chrome browser immediately after a web page loaded, without any additional clicks from the user.
In the Chrome browser, the model used four different flaws in sequence to execute commands outside of the browser's secure isolation environment. Google corrected these issues in Chrome version 156, which was released on October 20.
Regarding Safari, the identified flaw allowed malicious websites to bypass security policies to read data from other pages opened by the user. Apple corrected this flaw at the end of June.
For the Windows and Linux operating systems, the flaws involved gaps in the network stack that could cause server crashes or allow remote root access, which is total control over the affected machine. Corrections were distributed by developers and communities throughout the year, including a Linux kernel update identified as CVE-2026-98432.
Anthropic did not reveal which model was used or detail the specific steps of the process, making it difficult to determine the extent of human work involved. The company stated that the research aims to strengthen system defenses, noting that the same capability used by security teams to find and fix vulnerabilities could help a malicious agent exploit them.
Anthropic has expanded its security verification program to allow government agencies and company teams to use advanced versions of the technology to audit codes and anticipate flaws before cybercriminals can act.