THE ESSENTIALS IN 30 SECONDS

  • GTIG found exploit_agent on an exposed staging server: it compares code before and after patches, drafts exploits and runs them automatically, a first for the group's investigations.
  • A suspected North Korean cluster, MIDNIGHT NEPTUNE, edited .claude/settings.json in victims' repositories so that launching Claude Code downloaded the SOMBERMEME backdoor.
  • Google counts an average of 11 zero-day exploits per month in 2026, up from 8 in 2025, with a spike to 22 in August.

Google Threat Intelligence Group (GTIG) published its latest AI Threat Tracker on November 16, written with Mandiant. Its central claim is that attackers have moved from asking chatbots for help to running multi-agent systems that carry out parts of intrusions on their own.

A pipeline from patch to exploit

The most notable finding is a framework its operator called exploit_agent, discovered on an exposed staging server. According to GTIG, it continuously compares versions of popular web application software before and after security fixes, works out which weakness a patch closes, writes an exploit and executes it without anyone approving each step. GTIG says this is the first time its investigators have seen vulnerability discovery and exploitation fully automated by a threat actor.

The same server held stolen victim data and scripts for internal reconnaissance, brute-forcing, installing webshells and even removing webshells left by rival attackers. Victim telemetry pointed to French media, entertainment and education websites. Configuration comments were written in Russian, but Google does not attribute the framework to a named group, and Hackread notes that the report gives no success rate.

Open-source agents and a Claude Code backdoor

GTIG also describes financially motivated and suspected China-linked actors using open-source agentic pentesting frameworks. One called CAIRN appeared in intrusions against targets including the US aerospace sector, while a customised platform built on PentAGI was used for scanning organisations in Asia between late August and early October. Google says these tools currently favour speed over stealth and leave recognisable traces.

A suspected North Korean cluster tracked as MIDNIGHT NEPTUNE targeted cryptocurrency firms in North America and Europe through fake recruiter profiles and "ClickFix" lures. After gaining access, it modified the .claude/settings.json file in internal code repositories so that starting Claude Code triggered the download of a backdoor named SOMBERMEME.

Vulnerability trends and Google's response

The report says disclosed vulnerabilities and exploits nearly doubled between January and August 2026. Zero-day exploits averaged 11 per month this year against 8 in 2025, before reaching 22 in August. GTIG argues that language models make it cheaper to weaponise already-patched flaws, and notes a side effect: a stream of AI-generated proof-of-concept exploits that do not actually work.

Google says it disabled accounts and other assets linked to the misuse it observed, including activity involving Gemini. It does not say how many accounts were affected. As summarised by Hackread, the practical lesson for defenders is to patch internet-facing systems quickly and to look for signs of earlier compromise, since a fix does not remove an intruder who is already inside.