Google says Russian-speaking hackers ran the first fully automated AI cyberattack
Google said Monday that Russian-speaking hackers carried out what its analysts believe is the first cyberattack run end-to-end by artificial intelligence, without a human in the loop at any stage of the intrusion.
The operation appears in a quarterly report from Google’s Threat Intelligence Group and its Mandiant incident response unit. Researchers stopped short of naming a group or state.
Their attribution rests on Russian-language comments in the operation’s code, including a configuration file instructing its user to enter API keys for “auto-exploitation” there rather than in a chat.
Victim data on the staging server pointed to targeting of French websites in the media, entertainment and education sectors.
“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.
The centerpiece is a system the researchers dubbed “exploit_agent.” It chains together multiple AI agents with frontier and locally hosted language models to continuously compare code in widely used software before and after vendors publish patches. When a fix appears, the system infers what the change reveals, drafts an exploit and executes it automatically.
Investigators also found scripts for internal reconnaissance, password guessing and installing webshells, including one that erased the webshells of rival attackers. Google said it had identified threat actors using such tools “to fully automate vulnerability discovery and exploitation without a human in the loop for the first time.”
Until now, Google’s trackers described adversaries using AI mainly to accelerate work that people still directed: composing phishing lures, writing code, summarizing stolen data. The new report marks what the company calls a critical inflection point, with the pipeline from vulnerability research to intrusion running unattended.
Google also said it observed open-source offensive frameworks, including CAIRN and PentAGI, in activity targeting an aerospace organization in the United States and organizations across Asia. It also cited a China-nexus group building an automated attack framework for likely espionage work.
Speed is the clearest advantage. After vendors patched two PHP vulnerabilities on Oct. 21, hundreds of purported exploits appeared within days, most of them nonfunctional and likely AI-generated. Google counted 22 exploited zero-days, flaws unknown to vendors, in August alone, up from 11 per month earlier in 2026 and an average of eight in 2025.
The findings arrive amid mounting concern over autonomous AI systems, including Anthropic’s April release of Mythos, a cybersecurity-focused model, and a summer incident in which an OpenAI agent hacked development platform Hugging Face, though itdid not appear aimed at specific targets.
You can read the full report on Google’s website.