Skip to content
Advertisement
Google warns autonomous cyberattacks have arrived

Google warns autonomous cyberattacks have arrived

Google 2015 logo (Wikimedia Commons, public domain)

Google's threat intelligence team found an attacker framework that finds and exploits web vulnerabilities with no human in the loop

Google says the era of autonomous cyberattacks is no longer a conference-panel hypothetical. It has started.

In its Q4 2026 AI Threat Tracker, published November 16, 2026, Google’s Threat Intelligence Group (GTIG) describes what it calls the first fully automated vulnerability discovery and exploitation carried out by malicious actors.

Inside exploit_agent

The report, produced alongside Mandiant, centers on a compromised staging server belonging to an attacker. On that server, investigators found a multi-agent framework with a refreshingly unsubtle name: exploit_agent.

The framework connects on its own to a range of large language models. It then puts those models to work studying the code of popular web applications, both before and after security patches are applied.

Security researchers have used this technique for years. What GTIG flags as new is that exploit_agent runs the whole loop by itself, continuously, spotting weaknesses and writing working exploits without a human supervising.

Advertisement

GTIG also highlights a second worry: these automated operations can move fast while leaving relatively few traces behind.

Who got hit, and what was left behind

The campaign tied to the server mostly targeted French websites. The victims sat in three sectors: media, entertainment and education.

Investigators examining the server turned up stolen data, which confirms the framework was doing more than running practice drills. They also found post-compromise scripts, including brute-force tools and web management systems.

GTIG reads that setup as evidence of a sophisticated level of operational organization.

The report describes the operation as an advanced stage of attack automation.

From sidekick to operator

The shift becomes clearer when set against Google’s earlier findings. In GTIG’s Q3 2026 report, the team documented AI-assisted workflows among attackers but found no fully autonomous systems.

Earlier in 2026, AI largely played a supporting role in malicious operations. It helped with tasks like credential harvesting, acting as a productivity tool rather than the one running the show.

The report also lays out a timeline for what may come next. GTIG expects unsophisticated autonomous attacks to show up within one to three months, with more sophisticated versions expected within three to six months.

GTIG’s own conclusion points toward fighting automation with automation. The group says defenses will need to become advanced and automated to stay effective against these tactics.

Disclosure: This article was edited by Diego Almada Lopez. For more information on how we create and review content, see our Editorial Policy.