Anthropic says internal AI model built working Chrome, Safari, Windows, Linux exploits
An unreleased Claude model developed five working exploit chains, Anthropic said. The company says every flaw was reported to vendors in advance and has since been patched.
Symbolic image · Photo: Anthropic / Wikimedia Commons, CC0 (cropped and resized)
Anthropic's Frontier Red Team described in a research post five attack chains that an internal, unreleased AI model built in a lab setting. Targets included Chrome on macOS, Safari, a locked Android phone, Windows and Linux servers. The company said the work required only initial prompting and little human involvement. None of the attacks were observed in the wild.
In the Chrome attack, the model chained four flaws so that simply loading a crafted page ran code outside the browser sandbox, according to Anthropic. In total the company said it reported 62 Chrome flaws to Google, all rated critical or high, fixed in Chrome 156 released Oct. 20. On Linux, the model gained root access without login via a multipath TCP service; the Windows chain was a denial-of-service attack; and on Android it copied data off a locked device over USB.
Anthropic said all vendors were notified first and have shipped patches; Google and Apple credit Anthropic researchers working with Claude in their advisories. Security site The Hacker News noted, however, that some of the cited CVE records describe different bugs than Anthropic does. How much of each chain the model built on its own cannot be independently verified.