For years, security experts warned that AI would one day find and weaponize software bugs on its own. Google now says that day has arrived.

In its latest AI threat report, published this week, the Google Threat Intelligence Group describes what it calls the first fully autonomous case it has seen: an AI system that discovered a vulnerability, built a working exploit and launched it, with no human in the loop.

How it worked

The tool, which its operators named "exploit_agent," links several AI models together, both large commercial models and smaller ones run locally. It was set up to watch for security updates to popular web software, compare the code before and after each patch to spot what had been fixed, and then write and fire off attacks against systems that had not yet updated.

Google found it on an exposed staging server. The same server held data stolen from victims and scripts for scouting networks, guessing passwords and planting web shells. The operator appears to be Russian-speaking, and the targets included French media and entertainment sites and the education sector.

Not an isolated case

Google says it is tracking a growing list of similar AI attack frameworks, including ARTEX, PentAGI, HexStrike AI and others, used by both cybercriminals and suspected Chinese espionage groups. Because these tools are cheap, easy to update and able to scale, Google judges it highly likely that attackers will keep using them worldwide.

What it means

The gap between a patch being released and an attack using it has always been a race. When software can read a fix and turn it into a weapon automatically, that race gets shorter for everyone. For defenders, the lesson is plain: patch internet-facing web applications fast, watch for unusual behavior right after updates are published, and assume attackers are already reading your vendors' release notes.

Photo: Unsplash