Google Threat Intelligence Group (GTIG) has identified attackers using large language model (LLM) frameworks to automate software vulnerability discovery and exploitation without step-by-step human direction for the first time.
In a technical report published in November 2026, researchers detailed findings from an exposed staging server. This infrastructure hosted a multi-agent system comparing software versions to identify flaws and execute malicious code.
According to Google’s documentation, these automated workflows give organizations significantly less time to apply security patches before attackers compromise vulnerable systems.
The server ran a framework dubbed exploit_agent, configured to conduct continuous differential analysis before and after patch announcements on popular web applications.
Autonomously, the model drafted exploits and launched them against live targets. Telemetry on the server indicates the operators specifically targeted French websites across the media, entertainment, and education sectors.
Within the decentralized ecosystem, automated defense with AI agents gains relevance given the speed at which offensive tools operate against exposed code.
Shrinking Patching Windows
The November findings contrast with Google’s previous September 2026 report, when GTIG stated it had not yet observed fully autonomous pipelines deployed against targets in the wild.
The shift toward human-out-of-the-loop systems marks an operational inflection point in cybersecurity. Vulnerability disclosures more than doubled between January and August 2026.
The monthly average of exploited zero-day vulnerabilities rose from eight in 2025 to eleven between January and August 2026. During August alone, that figure reached a peak of 22 exploited flaws.
Alongside exploit_agent, researchers observed active deployment of open-source offensive tools including CAIRN and PentAGI in live intrusions targeting aerospace and Asian organizations.
These platforms chain complex attack paths at machine speed, though they currently trade stealth for velocity due to recognizable forensic traces left behind.
Parallel Risks in the Crypto Sector
Attack automation coincides with other campaigns specifically targeting the digital asset industry through artificial intelligence tools.
Previously, security researchers documented an AI-powered North Korean malware campaign utilizing advanced social engineering and deepfakes to compromise crypto startups and DeFi protocols.
Google Security Operations deployed specialized behavioral and volumetric rulepacks to detect automated offensive frameworks. Mandiant Consulting also released hardening guides to counter the operational speed of these tools.
GTIG’s quarterly update on November 16, 2026, confirms that autonomous offensive operations have moved beyond theoretical experimentation into live environments.

