A thief could have pulled photos off a locked Android phone, Anthropic's AI found
Big: fixes for most of these flaws are already out, so update your browser, phone and computer.
What happened
Anthropic, the company behind the Claude chatbot, says it used an unreleased, internal research model to find serious security holes in Chrome, Safari and Android, plus flaws in Windows and Linux that mainly affected servers, the computers that run websites. It says the AI also built working attacks. In Chrome, just opening a bad web page could have let an attacker run any program they chose on your computer. A thief holding a locked Android phone could have pulled photos from it over a USB cable. Anthropic says it reported the flaws, and the ones used in the five attacks it describes have been fixed.
What's actually going on
Software is full of small mistakes. Most are harmless, but a few let a stranger make a program do things it shouldn't. Browsers and phones are built in layers, so one mistake usually isn't enough. An attacker has to find several and link them, like getting past the front door, then the inner door, then the safe. Anthropic says its AI did that. In Chrome, it linked four flaws to escape the walled-off area where web pages are supposed to stay. In Safari, a bad site could have read your email if you were signed in.
Here's why the AI part matters. Anthropic says this work used to belong to a few expert teams, and now AI models like this one can do it "with little human involvement." Anthropic says it ran large numbers of AI agents (AI that works through a job step by step on its own) at the same time, each hunting in a different way. For Windows, people gave the AI the program file and tools, and it worked backward from the finished program to rebuild how it works, then spotted the flaw.
There are limits. People still aimed it: each attack began with what Anthropic calls "initial prompting," meaning a person gave the AI its starting instructions. The post doesn't say how many of the AI's leads were dead ends. Anthropic says Google has patched (fixed) 62 Chrome flaws from its reports, all rated critical or high severity.
The word you'll hear
Exploit
A piece of code that uses a flaw in software to make your phone or computer do something its makers never meant, like let a stranger read your data.
In your life
The fixes are already out, so the real risk now is running old software. If you keep tapping "later" on updates, this is a good week to stop. My guess: we'll see more stories like this, because if one AI company can find holes like these, others probably can too.
Be careful about
Anthropic says this skill used to belong to a few expert teams and now takes little human involvement. Attackers may get it too. And this is Anthropic describing its own unreleased model. The post doesn't say how often the AI got things wrong.
Where to get it
There's nothing to sign up for, since the model isn't released. The useful step is free: update. Anthropic says Chrome 156 includes its fixes. In Chrome, open the menu, then Help, then About Google Chrome. Then check for updates on your phone and computer.
Go deeper
Discovering and exploiting vulnerabilities in critical, real-world software with Claude
Anthropic, October 27, 2026
How was this brief?
Get one of these every morning
One story, about five minutes, in your inbox by 7 AM Eastern. No spam, and unsubscribing takes one tap.