Google says misconfigured test server exposed autonomous AI attack operation; system contained Russian-language comments
Google detects first AI-powered autonomous cyberattacks, linking them to Russia, China, Iran, and North Korea
Google's Threat Intelligence Group (GTIG) and Mandiant released a report saying that state-backed hacker groups have begun using networks of AI agents to carry out autonomous attacks, finding vulnerabilities, writing exploit code and breaching networks on their own. Google analysts found an internet-exposed test server whose exploit-agent system consisted of multiple agents connected to different LLMs, with notes in Russian; it automatically located vulnerabilities by comparing code differences before and after patches to applications such as WordPress plugins, generated exploits and attacked real targets, leaving stolen data on French media and education websites. The report said a China-linked group used AI to automate victim identification and attack tool development, North Korea's Midnight Neptune tampered with configuration files of coding assistants such as Claude Code to plant backdoors in intrusions targeting cryptocurrency companies in North America and Europe, and Iran-backed APT42 used Gemini for target profiling and social engineering.
The report documented real cases of state-linked groups using AI agents to autonomously find vulnerabilities and automatically launch attacks, and also noted that fully automated attacks currently trade stealth for speed.