On November 16, 2026, Google documented AI that exploits patches automatically: a tool called exploit_agent compares code before and after patches, writes exploits and runs them with no human in the loop. Google Threat Intelligence Group says zero-day exploits rose to 11 a month in 2026 and 22 in August.
Key facts
- Google Threat Intelligence Group and Mandiant reported on November 16, 2026 the first attackers seen using AI frameworks to find and exploit flaws with no human in the loop.
- The exploit_agent tool compares code before and after patch announcements, writes exploits and, per Google, targeted French websites in media, entertainment and education.
- Zero-day exploits rose from 8 a month (2025) to 11 a month (2026) and 22 in August, per GTIG; LLMJacking cost some firms over $100,000 in a single day.
Researched and written with AI, to Luigi's brief, positions and editorial rules. Automatically checked against the sources. How I work →
You know the moment: the plugin on your landing page asks for an update, and everyone thinks, next week. This week Google showed why that no longer works.
Google Threat Intelligence Group (GTIG) and Mandiant published their new AI Threat Tracker on November 16, 2026. The headline: for the first time, Google observed attackers using AI frameworks to find and exploit vulnerabilities fully automatically, with no human in the loop.
What Google found in the AI attack
Google calls the tool exploit_agent. It sat on an exposed staging server. According to the report, it works like this:
- It continuously compares the code of popular web applications before and after a patch announcement.
- It turns the difference into an exploit and runs it on its own.
- Scripts for internal reconnaissance, brute-forcing and installing webshells follow.
Google's telemetry suggests it targeted French websites in media, entertainment and education. One figure in the report shows the tool analyzing a WordPress plugin for the Gutenberg editor. The numbers show the pace: zero-day exploits rose from an average of 8 a month in 2025 to 11 a month in 2026, and 22 in August, per GTIG. For the PaperCut flaws, working exploits appeared within a day of disclosure.
Why this is a marketing problem
To me, this is the real point: the patch itself becomes the blueprint. The moment a vendor closes a hole, the code difference shows where it was. A machine reads that faster than your team schedules the update.
And who runs the landing pages, microsites and campaign pages with form plugins? Often marketing or an agency, not IT. That is exactly where plugins get installed once and forgotten. AI doesn't fix broken ways of working, it widens the cracks. Here the cracks are unmaintained pages with no owner. The Lovable study on apps with open databases showed how fast that happens.
The second part of the report matters just as much: stolen AI access, or "LLMJacking". Mandiant cites cases with more than $100,000 in fraudulent charges in a single day and up to $1 million in total. One route, per Google: keys from public GitHub repositories. That fits the Cursor breach, where keys in AI logs became master keys. That's why I set hard spending limits for I4U on every AI platform we use.
What I would do now
- Give every marketing website a named owner, including that 2023 microsite.
- Turn on automatic plugin updates where you can, and delete unused plugins instead of just deactivating them.
- Shut down old campaign pages nobody needs. Every open page is a door.
- Put a spending limit on every AI key your team uses, and check whether any of them sits somewhere public.
If you're building landing pages with AI or using AI agents in marketing, plan the maintenance from day one. And for how attackers use AI to deceive, see the North Korean deepfake meetings.
Bottom lineAn update is no longer a task for next week. It's a race, and the other side doesn't sleep.
Got AI news? Submit news →
FAQ
What is exploit_agent?
An attacker AI tool Google found on an exposed staging server. It compares code before and after patches, writes exploits and runs them with no human in the loop, per Google.
What is LLMJacking?
Stealing and abusing someone else's AI access. Mandiant cites cases with more than $100,000 in fraudulent charges in one day and up to $1 million in total.
How do I protect my marketing website?
Give every site an owner, turn on automatic plugin updates, delete unused plugins, shut down old campaign pages and put spending limits on AI keys.
Source
More sources
- SiliconANGLE – 16.11.2026, Hinweisgeber, Bericht zum GTIG-Tracker




