On Monday the FBI and AWS warned that AI agents were automating cybercrime at industrial scale. On Friday the other shoe dropped. The Register reports that attackers linked to China deployed AI agents that autonomously found and exploited a zero-day vulnerability, taking over a university server. Not a known bug from a list. A zero-day, discovered and weaponized by software acting on its own.
The detail that should stop you cold
According to Amazon's CISO, whose company has as much visibility into this kind of traffic as anyone on earth, the agents kept adapting for nearly a full day without human direction. Read that again. No operator guiding the loop, no human picking the next move. The agents probed, found a flaw nobody had catalogued, built the exploit and held the server for almost 24 hours, improvising the whole time. This is the Artex scenario the FBI described on Monday, except the humans are further out of the loop than the advisory suggested. 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 a university matters
A university server is not a Fortune 500 crown jewel, and that is precisely the point. Academic networks are soft targets, rich in research data and poorly defended, and they have been prime real estate for state-linked intrusions for a decade. The playbook is understood: get in quietly, establish persistence, watch. The new part is who did the getting in. When the entry work is delegated to autonomous agents, one operator can run fifty intrusions instead of five, and the bottleneck on state-scale espionage stops being skilled humans.
The week this landed in
Step back and look at the arc, because it is unnervingly tidy. Monday: FBI and AWS warn about industrial-scale AI agent crime. Wednesday: nginx, the web's most trusted load balancer, gets an unauthenticated remote code execution flaw plus a kernel bug that escapes containers. Thursday: the European Commission fights off an AI-assisted cyberattack while Europe's biggest gas supplier admits it has no recovery timeline. Friday morning: the White House fires its national cyber director after the outage cluster, with Trump saying he was not good enough. Friday afternoon: proof of concept for autonomous zero-days in the wild. Each story on its own has an explanation. Together they describe a defense establishment reorganizing in public while the threat it is chasing demonstrably outruns it. Source: https://www.bbc.co.uk/news/articles/cwz0k3nn5p5mo
What changes now
The vulnerability discovery pipeline just shifted. For thirty years the implicit assumption was that finding new bugs took rare human talent, which throttled how many zero-days existed. That throttle is gone. Expect the disclosed bug count, already past 100,000 this year, to keep climbing, and expect the window between discovery and mass exploitation to keep shrinking. The defenders' only real advantage left is the same technology on the other side, which is a cold kind of comfort, and it is exactly the world OpenAI's GPT-6.2 system card was drawn up for on Wednesday. The race between machine attackers and machine defenders started this week. The humans are spectators now, and one of them just lost his job over it.