Unreleased Anthropic AI Model Built Working Exploits for Chrome, Safari, Android, Windows and Linux
Anthropic says one of its unreleased AI models was able to discover security weaknesses and combine them into functional exploit chains affecting major software platforms, including Chrome, Safari, Android, Windows and Linux. The tests were carried out in controlled lab environments, and the company says all identified vulnerabilities were reported to the affected vendors before fixes were released.
According to Anthropic’s Frontier Red Team, the unnamed AI model created five separate attack chains with limited human assistance after receiving initial instructions. The company has not revealed which model was used, when it may become available, or exactly how much guidance researchers provided during each phase of the work.
One of the most striking demonstrations involved Google Chrome. The AI reportedly connected four separate vulnerabilities to break out of the browser’s security sandbox, a protection system designed to keep malicious web content isolated from the rest of the device. In the test, a malicious webpage loaded and, within seconds, the Calculator app opened on a Mac without any further action from the user.
Another test focused on Safari. In that case, the AI-developed exploit bypassed a security restriction that normally prevents websites from accessing data belonging to other websites. If used in a real-world scenario, such a flaw could potentially expose sensitive information from sites where a user was already signed in.
The Android demonstration showed how an attacker could extract photos and other data from a locked phone. However, the attack had an important limitation: it required physical access to the device through a USB connection. That makes it less practical for remote attackers, but still concerning in situations where a device could be seized or briefly accessed.
On Windows, the AI-generated attack targeted a server handling HTTP/3 requests. The exploit could repeatedly crash the server, creating a denial-of-service situation. Anthropic noted that this attack did not allow the attacker to run code or access stored files, but it could still disrupt services if used against vulnerable systems.
The Linux exploit was the most severe of the demonstrations. Anthropic says the attack could give a remote attacker full administrator-level control of a server without requiring a login or user interaction. However, the exploit depended on a specific network configuration, meaning it would not apply to every Linux machine.
Anthropic says the results show both the promise and the danger of advanced AI in cybersecurity. On one hand, AI systems could help security researchers identify complex vulnerabilities faster and report them before criminals can take advantage. On the other hand, the same capabilities could make it easier to combine isolated bugs into complete, real-world attack chains.
For now, these examples remain lab-tested demonstrations rather than evidence of active attacks in the wild. Still, the findings add to growing concerns about how powerful AI models may reshape vulnerability research, exploit development and digital defense. As AI tools become more capable, software companies and security teams may need to move even faster to find, patch and prevent critical flaws before they can be abused.
