There you have it! Anthropic's Frontier Red Team says an unreleased internal model found and expl…
October 28, 2026 · 0 likes · 0 comments
China Threat AI Cybersecurity
There you have it! Anthropic's Frontier Red Team says an unreleased internal model found and exploited flaws in Chrome, Safari, Android, Windows and Linux, and published five attack chains on Oct. 27.
The Chrome chain linked four bugs. Anthropic's proof of concept launches Calculator on a stock Chrome install on macOS within seconds of a page loading, no clicks needed. Anthropic says its reports led to 62 CVEs, all critical or high, all patched as of Chrome 156.
The Linux one is a zero-click remote code execution bug through multipath TCP. In Anthropic's test, about 48 minutes to leak a kernel pointer and about five more to get a root shell.
The caveats matter. Every flaw went to the vendor first, and the fixes are out. Google and Apple credits can be checked. What outsiders can't check is how much a human did at each step. Anthropic says the work started with "initial prompting" and needed little human involvement, but it hasn't published a step-by-step breakdown, and the model is unreleased.
Speaking only for myself, here's what keeps me up at night.
Anthropic found these, told Google, Apple and Microsoft, and waited for patches. That's how it's supposed to work.
So what happens when China has its own cyber-capable frontier models, finds the same class of zero-days, and never tells anyone?
No disclosure, no 90-day window, no patch. Just access.
This isn't a fantasy. Google's threat intelligence team attributed five zero-days exploited in the wild in 2024 to PRC-backed groups, and called them the most consistent government-backed espionage developer and user of zero-days. That's just the baseline we know about.
Defenders only patch what they know about. The whole patch cycle runs on someone choosing to disclose. Take that away and you're in a race you don't even know you're in.
And speed is the problem. The UBH story cites an Epoch AI chart: a couple dozen critical vulnerabilities reported in a month used to be normal. A typical month now sees about 600.
I've seen how slow our systems move. PHP2Shell showed us what a few days of lag looks like.
What I'd do today:
• Patch browsers and operating systems now, and shorten your patch window to days, if not hours, not months. Use agents to automate.
• Assume breach. Segment your networks so one foothold doesn't become everything.
• Scan your own code before it ships using agents with frontier models. Anthropic's advice is "scanning your own code is more important than ever." Agreed.
The good guys are disclosing. Plan for the ones who won't.
Full story, just the facts: https://lnkd.in/eAzPYc3V
The Chrome chain linked four bugs. Anthropic's proof of concept launches Calculator on a stock Chrome install on macOS within seconds of a page loading, no clicks needed. Anthropic says its reports led to 62 CVEs, all critical or high, all patched as of Chrome 156.
The Linux one is a zero-click remote code execution bug through multipath TCP. In Anthropic's test, about 48 minutes to leak a kernel pointer and about five more to get a root shell.
The caveats matter. Every flaw went to the vendor first, and the fixes are out. Google and Apple credits can be checked. What outsiders can't check is how much a human did at each step. Anthropic says the work started with "initial prompting" and needed little human involvement, but it hasn't published a step-by-step breakdown, and the model is unreleased.
Speaking only for myself, here's what keeps me up at night.
Anthropic found these, told Google, Apple and Microsoft, and waited for patches. That's how it's supposed to work.
So what happens when China has its own cyber-capable frontier models, finds the same class of zero-days, and never tells anyone?
No disclosure, no 90-day window, no patch. Just access.
This isn't a fantasy. Google's threat intelligence team attributed five zero-days exploited in the wild in 2024 to PRC-backed groups, and called them the most consistent government-backed espionage developer and user of zero-days. That's just the baseline we know about.
Defenders only patch what they know about. The whole patch cycle runs on someone choosing to disclose. Take that away and you're in a race you don't even know you're in.
And speed is the problem. The UBH story cites an Epoch AI chart: a couple dozen critical vulnerabilities reported in a month used to be normal. A typical month now sees about 600.
I've seen how slow our systems move. PHP2Shell showed us what a few days of lag looks like.
What I'd do today:
• Patch browsers and operating systems now, and shorten your patch window to days, if not hours, not months. Use agents to automate.
• Assume breach. Segment your networks so one foothold doesn't become everything.
• Scan your own code before it ships using agents with frontier models. Anthropic's advice is "scanning your own code is more important than ever." Agreed.
The good guys are disclosing. Plan for the ones who won't.
Full story, just the facts: https://lnkd.in/eAzPYc3V