Google finds attackers automating hacks with no human involved
A framework called exploit_agent compared software before and after patches, wrote exploits and ran them on its own, Google’s threat researchers said.
Summary

In brief
- A framework named exploit_agent studied patches, wrote exploits and ran them on its own
- Data on its server points to French media, entertainment and education websites
- Google expects attacks with little to no human interaction within one to three months
Attackers have for the first time used AI to find and exploit software vulnerabilities with no human in the loop, Google’s Threat Intelligence Group said in its AI Threat Tracker report on November 16.
The finding leaves defenders less time to install a fix once a patch is announced. The report says the adversarial use of AI has reached “a critical inflection point,” and its timeline places current activity at stage 5 of 7, where AI tools carry out several stages of an attack with minimal human involvement.
The group found the framework, named exploit_agent, on an exposed staging server. It connects to several frontier and locally run language models and continuously compares versions of popular web application software before and after patches are announced, a method called differential analysis. It then identifies likely vulnerabilities, drafts exploits, code that takes advantage of a flaw, and runs them automatically. One example in the report shows it inspecting access controls in Spectra, a WordPress plugin.
Data on the server suggests the operator targeted French websites in media, entertainment and education, the group said. Scripts stored alongside the framework handled follow-up steps such as brute-forcing passwords and installing webshells, hidden backdoors on a web server. Comments in a configuration file for the framework’s language-model keys were written in Russian.
Exploitation is speeding up more broadly. Exploits of known flaws nearly doubled between January and August 2026, the report says, while attacks on zero-days, flaws that have no fix yet, rose from an average of 8 a month in 2025 to 11 a month in 2026 and 22 in August.
For now, automated attack frameworks trade stealth for speed, Google said, generating network traffic and code-level traces that make them easier to detect. In its security forecast for 2027, published on November 10, the company said defensive AI agents will take over sorting and investigating alerts, “surfacing a verified conclusion rather than a raw ticket.” Sandra Joyce, vice president of Google Threat Intelligence, said in the forecast that adversaries are poisoning open-source software and libraries that AI coding assistants pull into code bases automatically.
Government agencies have been urging defenders toward the same tools. Australia’s signals intelligence agency told organizations in October to use advanced AI models for cyber defense as the incidents it handled rose 20%, and Google DeepMind gave cyber defenders first access to its Gemini 4 Carbon model on October 28.
Google’s timeline projects stage 6, multi-agent attacks with little to no human interaction, within one to three months, and fully autonomous attacks within three to six months.