Google Says AI Agents Have Run Their First Fully Autonomous Cyberattacks

Rows of server racks in a data centre. Photo: Fleshas / Wikimedia Commons (CC BY-SA 3.0)

Google’s threat researchers say they have watched artificial intelligence run a hacking campaign from start to finish without a human at the keyboard, the first time they have observed the full chain of attack automated in the wild.

The finding is the headline of the Google Threat Intelligence Group’s quarterly AI Threat Tracker, published on 16 November with Mandiant under the title “Agentic Threat Activity Crosses a Significant Threshold”. The researchers found a multi-agent framework they call exploit_agent on a server its operator had left exposed to the internet. The system compares a piece of popular web software before and after a vendor publishes a security patch, works out which flaw the fix closes, drafts an exploit for it and fires it at servers still running the old version — all without a person directing each step.

Telemetry from the staging server suggested the operator was focused on French websites in the media, entertainment and education sectors. Google printed one example of the framework at work: it inspecting access-control code in Spectra, a WordPress plugin with more than a million active installations. Google does not say the agent found a flaw in Spectra or exploited it, and it gives no count of sites compromised and no success rate for the framework.

The report places exploit_agent in a wider shift. Google says it has now seen several threat actors abusing open-source, multi-agent penetration-testing frameworks — including ARTEX, CAIRN, CyberStrikeAI and PentAGI — to automate stages of live intrusions. CAIRN was observed against targets including the US aerospace sector, and a customised PentAGI platform scanned organisations across Asia between late August and early October. Vulnerability disclosures nearly doubled between January and August, from 5,045 to 10,740, while zero-day exploitation averaged 11 a month in 2026 against eight a month in 2025, spiking to 22 in August. The report also describes “LLMJacking”, in which criminals hijack cloud accounts to burn victims’ computing and AI-service quotas.

Google names no group or country behind exploit_agent, and the attribution remains unverified: some reports have pointed to Russian-language comments in the operation’s code, but Google itself does not make that claim. The practical warning is about speed. Because the automation targets flaws that already have a patch — so-called n-days — the window between a fix being published and an exploit arriving is collapsing. Google projects attacks needing little human involvement within one to three months and fully autonomous attacks within three to six months, though those are forecasts, not observations.

Sources: Google Threat Intelligence Group and Mandiant, “GTIG AI Threat Tracker: Agentic Threat Activity Crosses a Significant Threshold”, Google Cloud Blog, 16 November 2026. https://cloud.google.com/blog/topics/threat-intelligence/agentic-threat-activity-crosses-a-significant-threshold

Leave a Reply

Your email address will not be published. Required fields are marked *