Wednesday 18 November 2026 Edition compiled at 07.45 NL EN FR
← Tech · Other tech
Tech · Other techCybersecurity

From security patch to working break-in without a human stepping in: Google uncovers an attack system built from AI agents

According to Google and Mandiant, Russian-speaking hackers used it to target French websites, and the pace at which flaws are exploited has doubled.

· latest source update ma · 2 sources · CyberScoop, Business Standard
Illustration for this article

Google says it has for the first time seen an attack in which an AI system found and exploited vulnerabilities entirely on its own. That is in the quarterly report of the Google Threat Intelligence Group and incident response firm Mandiant, published on Monday, CyberScoop reports. It concerns a system called exploit_agent, which the researchers found on an exposed server.

The system links several AI agents together, using both well-known commercial language models and models running locally. According to Business Standard, it continuously tracks updates to widely used software, compares the code before and after a security patch, infers where the flaw was and then builds and runs an attack itself. The server also held scripts to look around inside a network, guess passwords and install backdoors. One script even wiped the backdoors of rival attackers, CyberScoop writes.

Google does not name a specific group, but Russian-language comments in the code point to Russian-speaking perpetrators. Victim data on the server pointed to French websites in media, entertainment and education. Google also saw open attack frameworks such as PentAGI and CAIRN in use against organisations in Asia and a US aerospace company, as well as a China-linked group building its own automated system, probably for espionage.

The report also shows how fast things move: in August Google counted 22 exploited zero-days, compared with an average of eight per month in 2025. After two PHP patches on 21 October, hundreds of supposed exploits appeared within days, most of them non-functional and probably written by AI. Google does note that such automated attacks favour speed over stealth and leave clear traces, which makes detection easier.

Spotted a mistake in this article? Let us know.