This week AI crossed a threshold, and it cuts both ways. The Register reports that attackers linked to China deployed AI agents which autonomously found and exploited a zero-day vulnerability, taking over a university server. No human found that bug first. But here is the part worth holding onto: the same capability that finds a zero-day in the wild can find and fix the next thousand before anyone gets hurt. One side proved it works. The other side just got handed the same machine.

What actually happened

According to Amazon's CISO, whose company sees more of this traffic than anyone, the agents kept adapting for nearly a full day without human direction. They probed, found a flaw nobody had catalogued, built the exploit and held the server, improvising the whole time. That is a genuinely new capability arriving in the world, and it belongs to whoever points it at a problem. Source: https://www.theregister.com/security/2026/11/20/china-attackers-deployed-ai-agents-that-autonomously-found-and-exploited-a-0-day-to-take-over-a-university-server/5311382

Why this is good news for defense

For thirty years, finding new bugs took rare human talent, and fixing them took even longer. The backlog is the story of modern computing: a hundred thousand disclosed vulnerabilities this year alone, and that is just what humans managed to catch. Agents that can discover zero-days can also be aimed at the fix list, and they do not get tired, do not miss the obvious, and work at the speed of the exploiters. Debian is already trimming APT features because AI-assisted research keeps surfacing holes faster than humans can audit them. That is the direction everything is about to go: from humans patching code they wrote years ago, to machines auditing code continuously while it is still fresh. Source: https://www.theregister.com/security/2026/11/20/china-attackers-deployed-ai-agents-that-autonomously-found-and-exploited-a-0-day-to-take-over-a-university-server/5311382

The week that proved the point

Look at the arc. Monday, the FBI and AWS warned about industrial-scale AI agent crime, which is really a warning that defense has to automate too. Wednesday, nginx got an unauthenticated RCE, and the fix shipped the same day. Thursday, the European Commission fought off an AI-assisted attack, and Friday the White House changed its cyber leadership while saying out loud that advanced AI models can find vulnerabilities humans cannot. Every one of those stories has the same shape: a faster offense forcing a faster defense, and the defense catching up in public. Source: https://www.bbc.co.uk/news/articles/cwz0k3nn5p5mo

What to do with it

If you run infrastructure, this is the week to point AI at your own code before someone else's AI points itself there. The tools that autonomously exploit are the same tools that autonomously audit, patch and verify. The teams that adopt them first will spend their time on work that matters instead of drowning in a bug backlog humans could never clear. The zero-day was going to be found eventually; it always was. What changed this week is that we now have machines that can find the rest of them too, and fix them, at the same speed they get broken. That is the story.