
Anthropic has revealed that an unreleased AI model built five exploit chains targeting widely used software platforms. The model identified and combined vulnerabilities across Chrome, Safari, Android, Windows, and Linux. This situation shows how advanced AI can turn separate software weaknesses into multi-stage attacks.
Anthropic Reveals Five Exploit Chains Across Major Platforms
During controlled testing, the model helped develop five attack chains. In Chrome, it combined four vulnerabilities to escape the browser sandbox and execute code on a Mac. A proof of concept launched the Calculator app seconds after a malicious page loaded, without further user interaction.
The Safari chain bypassed the same-origin policy, which normally prevents websites from accessing other sites’ data. Because of this, the exploit could expose browsing history, cookies, and information from authenticated sessions. On Android, three flaws enabled data extraction from a locked Google Pixel 9 through a physical USB connection.
Additionally, the Windows chain targeted the kernel’s QUIC handling and could remotely crash systems processing HTTP/3 traffic. However, the demonstration showed denial of service, not remote code execution. In contrast, the Linux chain targeted servers running a multipath TCP listener. The model found a kernel address within 48 minutes and reached root command execution five minutes later.
Following these demonstrations, Anthropic reported the vulnerabilities to affected vendors through coordinated disclosure. Google assigned 62 CVEs to Chrome vulnerabilities and addressed them in Chrome 156. Also, Apple and Google addressed the reported Safari and Android issues while vendors fixed the Windows and Linux vulnerabilities.
AI-Driven Exploits Raise New Cybersecurity Concerns
The research examined whether advanced models could identify separate weaknesses and combine them into functional exploits. Before now, specialists needed considerable technical expertise to develop multi-stage attacks. According to Anthropic, the model completed much of the work after initial prompting and also used parallel agents to examine software for further weaknesses.
Consequently, advancement like this could change the economics of vulnerability research. When AI automates discovery and exploit development, attackers may need less time and specialist knowledge. More importantly, combining flaws can defeat several protections in sequence, turning limited weaknesses into wider compromises.
In addition, this shift complicates patch management because one exposed component can affect products and services. Therefore, security teams need visibility across browsers, operating systems, network services, and application dependencies before attackers connect weaknesses.
Nevertheless, the model’s autonomy remains unclear. Anthropic has not named the unreleased system or detailed how much human involvement each exploit required.
Defenders Prepare for More Capable AI Systems
To reduce these risks, Anthropic wants security teams to use advanced models for defensive research. The company recommends continuous code scanning, including during pull-request reviews. It has also published security guidance for Claude Code to help developers identify vulnerabilities.
Alongside these recommendations, Anthropic expanded its Cyber Verification Program in October. The program gives vetted cybersecurity professionals access to advanced models for authorized security work. These activities include incident response, malware analysis, and penetration testing.
Ultimately, vendors must identify and patch vulnerabilities quickly, while organizations must install updates and review exposed services. As AI models advance, defensive access, careful oversight, and timely patching will shape cybersecurity preparedness across connected systems.
