Category: AI Agent Security | Reading time: 3 min
Google's Threat Intelligence Group (GTIG) says agentic threat activity has crossed a significant threshold. In its AI Threat Tracker published November 16, the team describes a criminal multi-agent system called "exploit_agent" that found vulnerabilities, wrote exploits and launched them without a human in the loop. GTIG calls it the first time it has seen discovery-to-exploitation fully automated in the wild.
How exploit_agent worked
GTIG found the tooling on an exposed staging server. The setup was wired to multiple frontier and locally hosted language models. It ran continuous differential analysis on popular web application software, comparing code before and after a security patch to work out which flaw the patch closed. It then drafted an exploit and executed it automatically.
One example in the report is an access check in version 2.19.25 of the Spectra (Ultimate Addons for Gutenberg) WordPress plugin. Configuration comments on the server were written in Russian and referred to an "exploit_agent pilot."
The same server held exfiltrated victim data and post-compromise scripts for reconnaissance, brute forcing, installing webshells and removing other attackers' webshells. Victims were mostly French media, entertainment and education sites.
The wider picture
GTIG says attackers are also using the ARTEX, CAIRN, CyberStrikeAI and PentAGI agent frameworks in live intrusions. The report puts zero-day exploitation at an average of 11 per month in 2026, up from 8 per month in 2025, with a spike to 22 in August 2026.
It also describes LLMJacking, where stolen model-provider credentials are resold through "LLM Mixers," and espionage groups tampering with coding-assistant configurations to plant backdoors.
Why it matters for agent deployers
The patch window is the time between a fix shipping and a company installing it. When an agent can read the fix and attack within that window automatically, the window gets shorter for everyone. The report also shows the same agent frameworks defenders evaluate being used on the offense side.
For enterprises running their own agents, the lesson runs both ways. Attackers are now operating at machine speed, and internal agents with broad credentials become higher-value targets. Knowing exactly what each agent is allowed to do, and being able to show what it actually did, becomes part of basic incident response.
Source: Google Threat Intelligence Group: Agentic threat activity crosses a significant threshold | SecurityWeek | BleepingComputer