Anthropic says internal AI model built working exploit chains

Anthropic says an unreleased internal AI model found security flaws and combined them into working proof-of-concept exploits across browsers and operating systems. The company described five chains built in laboratory tests after prompting; all the flaws have been patched, and the exploits were not observed in real attacks. One Chrome demonstration opened Calculator on a stock Mac after a page loaded, without further clicks. A Linux chain obtained a root shell over the network but required a server running a multipath TCP listener. Apple and Google credit Anthropic and Claude for specific fixes. Those acknowledgments do not establish how independently the model worked: Anthropic has not broken down human involvement at each step.