Skip to content
ShiftDelete.Net
Search news/ TR Türkçe
News Feed AI Smartphones Games Cars Computers Social Media Space Reviews Latest news
  1. Home
  2. News

Google Catches Hackers Using AI to Turn Patches Into Exploits With No Human Help

A new Google Threat Intelligence report says an attacker's "exploit_agent" system compared patched code, wrote exploits and launched them on its own, while stolen AI accounts now sell for as little as US$0.10 per million tokens.

Hakkı ALKANHakkı ALKAN

Published 5 min read

Empty security operations desk with monitors showing network graphs, illustrating AI cyberattacks

Google says attackers have crossed a line security researchers have been warning about for two years. In a report published on 16 November 2026, the Google Threat Intelligence Group (GTIG) and Mandiant say they caught threat actors using large language model (LLM) frameworks to find software flaws and exploit them with no human in the loop, the first time Google has seen that end to end.

The finding sits at the top of the latest GTIG AI Threat Tracker, which also covers stolen AI accounts being resold by the million tokens, a North Korean group hiding a backdoor in Claude Code settings, and open-source "agentic" hacking kits now showing up in real intrusions.

An exposed server running "exploit_agent"

GTIG found an exposed staging server hosting a multi-agent framework its operator called exploit_agent. It plugs into several frontier and locally run LLMs and watches popular web application software around the clock.

When a patch is announced, the framework compares the code before and after the fix, works out which weakness was closed, drafts an exploit and runs it against targets automatically. The same server held data already stolen from victims, plus scripts for the next steps after a break-in: internal reconnaissance, password brute-forcing, installing web shells and even deleting web shells left by rival attackers.

According to Google, victim data on the server suggests the operator was going after French websites in media and entertainment, as well as the education sector. A configuration file carried Russian-language comments describing the setup as an "exploit_agent pilot". Google has not tied the activity to a named group.

One figure in the report shows the framework inspecting access-control code in version 2.19.25 of the Spectra (Ultimate Addons for Gutenberg) WordPress plugin. Google does not name a CVE for the plugin or say an attack on it succeeded. We checked the WordPress.org plugin directory on 16 November: the plugin, now listed as Spectra Legacy, has more than 1,000,000 active installs and its current version is 2.20.6, last updated on 3 November 2026. That is no evidence of a live flaw, but it shows how quickly an outdated release can end up in an attacker's test bench.

Why patched bugs are the new target

The report explains why attackers are pointing AI at "n-day" bugs, flaws that already have a fix, rather than hunting for unknown zero-days. A patch narrows the search: an AI agent can compare source code versions, including preview builds, read release notes and study proof-of-concept code, then turn that into a working attack before many organizations have installed the update.

Google gives two recent examples. In late August 2026, two zero-days in PaperCut print management servers (CVE-2026-81578 and CVE-2026-82078) could be chained for remote code execution without logging in. Researchers and security firm Huntress built working exploits with LLMs within a day of disclosure, and GreyNoise later reported an attacker who likely used LLMs to build exploits, pick targets and hit hundreds of organizations four days after disclosure, reaching domain administrator within minutes.

The PHP flaws CVE-2026-109618 and CVE-2026-109619 show the other side. They were found with LLM-assisted analysis, early patched code leaked before the official fix on 21 October 2026, and hundreds of fake, likely AI-generated proof-of-concept exploits followed. Defenders had to check each one.

AI cyberattacks by the numbers

GTIG says overall vulnerability disclosures and exploits nearly doubled between January and August 2026. Zero-day exploitation grew more slowly, from an average of 8 a month in 2025 to 11 a month in 2026, before jumping to 22 in August.

FindingWhat Google reports
Zero-day exploits per month8 (2025 average), 11 (2026 average), 22 (August 2026)
Worst single-day LLMjacking billMore than US$100,000
Worst cumulative LLMjacking abuseAbout US$1 million before cleanup
Stolen AI access resale priceUS$0.10 to US$25 per million tokens
Agentic hacking kits seen in the wildCAIRN and PentAGI (ARTEX, CyberStrikeAI and HexStrike AI also tracked)

Hacking kits that run themselves

Beyond exploit_agent, GTIG tracks a set of open-source "agentic" penetration-testing frameworks that attackers now use to automate parts of real intrusions. They include ARTEX, the tool whose control panels were found exposed online earlier this month, along with CAIRN, CyberStrikeAI, PentAGI and HexStrike AI.

Google says CAIRN was used against targets including the US aerospace sector, and a customized PentAGI platform scanned a wide range of organizations in Asia between late August and early October 2026. A China-linked group was also building its own automated attack framework. For now, GTIG says these tools "trade stealth for velocity": they are fast but noisy and leave clear traces. Google expects them to become both more autonomous and harder to spot.

Stolen AI accounts and poisoned coding tools

The report also describes "LLMjacking", in which criminals steal cloud and AI credentials and burn through the victim's model and GPU quotas. In one case Mandiant investigated, a botnet or proxy service made millions of LLM API calls in about 24 hours before the credentials were revoked. Stolen access is then resold through what Google calls "LLM Mixers" on Russian- and Chinese-language markets.

A suspected North Korean cluster Google tracks as MIDNIGHT NEPTUNE went after cryptocurrency companies in North America and Europe with fake recruiter profiles and "ClickFix" lures. Once inside, it edited the .claude/settings.json file in internal code repositories so that launching Claude Code would download a backdoor called SOMBERMEME, a reminder of the risks we covered in an earlier Claude Code extension attack.

What Google is doing and what you should do

Google says its Security Operations product now ships detection rules for ARTEX, Mandiant has published a hardening guide, and it has disabled accounts and assets linked to the misuse it found. It also points to its own defensive AI agents, Big Sleep and CodeMender, which hunt for and fix bugs, an approach rivals are also taking with tools such as OpenAI's Daybreak.

For website owners and IT teams, the lesson is simple: the time between a patch and an attack is shrinking. Install security updates for internet-facing software and plugins as soon as possible, keep staging servers off the public internet, and watch cloud AI bills and API keys for sudden spikes.