Anthropic's unreleased model chained flaws into working exploits across major browsers and operating systems
None were used in real attacks, and all were reported to vendors and patched, but outside readers cannot check how much of the work the model did alone.
Anthropic said an unreleased internal model found flaws in Chrome, Safari, Android, Windows and Linux and combined them into proof-of-concept exploits, none of which were used in real attacks and all of which were reported to vendors first. Its Linux chain is linked by Anthropic to a kernel use-after-free rated 9.8, though the public CVE record describes a different TCP fix, so that link rests only on Anthropic's account. Apple credits Anthropic and Claude for a Safari fix in iOS and iPadOS 26.5.2, and Google's Chrome release notes credit Anthropic-assisted work for some fixes, while Anthropic says it reported 62 Chrome flaws, not all of which are credited publicly. The Windows chain is a denial-of-service issue whose public record names HTTP.sys and credits no finder, and Anthropic says the Android flaw is fixed without a public record. Readers cannot independently verify how much of each exploit the model built without human direction, since Anthropic does not break that down.
Why it matters
The findings show that an AI model can connect several separate bugs into a working attack path, which raises the bar for how fast vendors must find and fix flaws. The Linux link and the share of work done without human direction rest on Anthropic's own account, so the scale of the risk cannot be confirmed independently. Disclosure rules that give developers 90 days to patch before details are discussed will shape how much of this work reaches the public.
Matters or hype?
Does this story matter, or is it hype? Vote before you see what everyone else said.