Skip to main content
Emerging ThreatsMalware & Ransomware

Google Exposes First Fully Automated AI Cyberattack by Russian Hackers

Rows of computer servers and networking equipment with a single terminal in the foreground displaying a blurred screen.

"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," Google's Threat Intelligence Group and its Mandiant incident response unit wrote in their quarterly report.

What Google’s investigators say they saw

Google’s teams report what they believe is the first end-to-end cyberattack run entirely by artificial intelligence, without a human in the loop at any stage of the intrusion. The operation—described in a quarterly report from Google’s Threat Intelligence Group and Mandiant—centered on a system the researchers dubbed "exploit_agent." According to the report, exploit_agent chains together multiple AI agents and uses both frontier and locally hosted language models to compare widely used software before and after vendors publish patches, infer the implication of changes, draft an exploit, and execute it automatically.

How the automated pipeline operated

Investigators said the pipeline included more than exploit drafting and execution. They found scripts for internal reconnaissance, password guessing, and installing webshells—one of which was built to erase webshells left by rival attackers. A configuration file in the operation's code instructed users to enter API keys for "auto-exploitation" there rather than in a chat, a detail Google flagged as part of the attribution to Russian-speaking operators based on Russian-language comments found in the code. The researchers stopped short of naming a specific group or state.

Speed and scale: numbers that changed during the campaign

Speed was a defining advantage, the report notes. After vendors patched two PHP vulnerabilities on Oct. 21, hundreds of purported exploits appeared within days; Google judged most of those to be nonfunctional and likely AI-generated. The company also recorded a jump in exploited zero-days: 22 exploited zero-days in August alone, compared with 11 per month earlier in 2026 and an average of eight per month in 2025. Google said it had identified threat actors using tools like exploit_agent "to fully automate vulnerability discovery and exploitation without a human in the loop for the first time."

Targets observed: France, the United States, and Asia

Victim data on a staging server analyzed by Google pointed to targeting of French websites in the media, entertainment and education sectors. Separately, Google reported observing open-source offensive frameworks—named in the report as CAIRN and PentAGI—in activity that targeted an aerospace organization in the United States and organizations across Asia. The report also cited a "China-nexus group" building an automated attack framework that the researchers characterized as likely intended for espionage work.

What this means for technologists, policymakers, and affected enterprises

  • Technologists and security teams: The report documents pipelines that remove humans from the vulnerability-to-exploit cycle, enlarging the window in which automated tools can discover and weaponize patches. Teams will be watching for automated exploit attempts, rapid surges of purported exploits after vendor fixes, and signs of automated webshell deployment and removal by competing attackers.
  • Policymakers and regulators: Google’s findings—particularly the rise in exploited zero-days and the public presence of offensive frameworks such as CAIRN and PentAGI—underscore arguments in the report about a changing operational baseline for cyber operations and the "critical inflection point" that automated exploitation represents.
  • Affected enterprises and procurement leaders: The observed targets—French media, entertainment and education sites, an aerospace organization in the U.S., and organizations across Asia—illustrate both sectoral and geographic diversity in targeting. The report’s details on rapid exploit proliferation after vendor patches highlight the operational pressure on patch management and incident response processes.

Google framed the findings alongside other recent developments it views as part of a larger trend: Anthropic’s April release of Mythos, described as a cybersecurity-focused model, and a summer incident in which an OpenAI agent accessed Hugging Face in an automated fashion (an event that Google’s report noted did not appear aimed at specific targets). The company concluded that the combination of low barriers to entry, ease of integrating new exploits, and massive scaling potential makes continued use of such frameworks by threat actors highly likely.

The full report is available from Google; read the CyberScoop summary and the underlying report at the link below.

Original story on CyberScoop