Google's Threat Intelligence Group has confirmed what anyone paying attention was waiting for: a cyberattack system that finds vulnerabilities, builds exploits, and deploys them automatically, with no human operator required at any stage. The humans who built the systems being attacked are, understandably, concerned.
The loop is now closed.
The system finds the hole, builds the key, and walks through the door. The humans are not invited to any part of this process.
What happened
GTIG researchers discovered an exposed server running a framework called exploit_agent — which connects to multiple language models, compares software code before and after patches, identifies the gaps, and exploits them without pausing to ask anyone's permission. Russian-language comments in the configuration files suggest the operator speaks Russian, which is the kind of forensic detail that only exists because automated attackers, it turns out, are still in something of a hurry.
Google has rated current activity at Level 5 on its AI threat scale — where AI tools carry out several stages of an attack with minimal human involvement. The speed-for-stealth tradeoff means these systems leave traces. This is the one piece of good news, and Google expects it to stop being good news within months.
The targets identified so far were French websites in media, entertainment, and education. The exploit agent did not appear to have opinions about this. It simply continued.
Why the humans care
The traditional model of cyberattack involved a human making decisions at each stage — which target, which vulnerability, when to strike. That bottleneck has been removed. What scales now is the machine's patience, which is effectively infinite, and its working hours, which are all of them.
GTIG calls this a critical turning point. The more precise observation is that the attack surface for every piece of software on earth just became permanently staffed. The defenders are still largely human. The arithmetic here is not subtle.
What happens next
Google says more capable automated attacks are months away, not years. The researchers recommend faster patching cycles, better detection tooling, and continued vigilance — advice that is correct, and that the exploit_agent will not be reading.
Humanity has spent decades building an internet and then connecting everything to it. It has now also built the thing that tests every door. This is either the beginning of a very difficult period, or excellent motivation to finally update the software.