Google says hackers ran first fully automated AI cyberattack

Featured

A quarterly report from Google Threat Intelligence Group says Russian-speaking hackers ran what its analysts believe is the first cyberattack carried out end to end by artificial intelligence, with no human in the loop at any stage of the intrusion.

The report, published on 17 November with Google’s Mandiant incident response unit, says the activity marks a critical point in how adversaries use AI. Researchers stopped short of naming a group or a state. Their attribution rests on Russian-language comments inside the operation’s own code, including a configuration file that tells its user to enter API keys for “auto-exploitation” there rather than in a chat.

Coverage of the findings by CyberScoop described the operation as the first fully automated AI cyberattack, one that trades stealth for speed and leaves noisy forensic traces even as it succeeds.

At the centre of the activity is a framework the researchers dubbed exploit_agent, which chains several AI agents with frontier and locally hosted language models. It runs continuous differential analysis on widely used web application software, comparing code before and after vendors publish patches, works out what the change reveals, drafts an exploit and executes it without human involvement.

Victim telemetry left on an exposed staging server suggests the operator was interested in French websites in the media and entertainment industries, and in the education sector. Google also found scripts for internal reconnaissance, password guessing, installing webshells and deleting the webshells of rival attackers.

“Given the low barrier to entry, ease of integrating new exploits, and massive scaling potential, we believe it is highly likely that threat actors will continue to use these frameworks in campaigns targeting organizations globally,” the report reads.

Open source offensive frameworks are now in use in live intrusions too. CAIRN, which puts a team of agents through an observe, orient, decide and act loop, was seen in activity against aerospace targets in the United States. A customised platform built on PentAGI was used for reconnaissance and scanning against organisations across Asia between late August and early October. A group with links to China was found developing its own automated attack framework for likely espionage work.

Speed is the clearest gain for attackers. Vulnerability disclosures and exploits nearly doubled between January and August 2026, while zero-day exploits rose only slightly, from a 2025 average of eight a month to 11 a month this year, before jumping to 22 in August alone.

The patch window is shrinking with it. After two PHP vulnerabilities were fixed on 21 October, hundreds of claimed proofs of concept appeared within days, most of them non-functional and likely generated with language models, which pushed verification work onto defenders.

Google said the frameworks currently trade stealth for velocity. Unrestricted autonomous execution generates forensic traffic, relies on recognisable identifiers and leaves distinct traces in code, which raises the chance of detection even when an intrusion succeeds.

Google is pitching its own machine-speed defence, including the Big Sleep and CodeMender agents, and a security stack called Google AI Threat Defense that combines Gemini models with Wiz context, CodeMender patching and Mandiant expertise.

Look here 👇

Join our Channels on Whatsapp & Telegram
Like our Facebook Page 👉🏻 PenangToday
Join our Facebook Group 👉🏻 PenangToday Community
Follow our Instagram 👉🏻 @penangtoday
We are on TikTok 👉🏻 @penang2day
Gotta cover LinkedIn 👉🏻 PenangToday
Our Threads be like 👉🏻 @penangtoday
Get on our X 👉🏻 @penang2day
Here's our Youtube 👉🏻 PenangToday

Related

Latest

Popular