Illustration of a browser window with a blue bandage and a purple magnifying glass examining it, with a magenta lightning bolt in the lens

Google: Attackers Use AI to Turn Patches Into Exploits

A
Admin
·

On 16 November 2026, Google's Threat Intelligence Group (GTIG) published a report saying it has, for the first time, seen attackers use AI to find and exploit vulnerabilities with no human in the process. One of those systems studies each web software patch in order to hit sites that have not installed it yet.

What we know

  • The report: "GTIG AI Threat Tracker", published by Google Threat Intelligence Group on the Google Cloud blog on 16 November 2026. The Hacker News covered it the same day.
  • The core claim: according to GTIG, in the past quarter it identified threat actors using LLM-based frameworks to fully automate vulnerability discovery and exploitation without a human in the loop, for the first time.
  • How the main case works: GTIG found an exposed staging server running a multi-agent framework, dubbed "exploit_agent", that continuously analyzes popular web application software. It compares code before and after each patch, works out what was fixed, drafts an exploit and runs it without human intervention. Telemetry on the server suggests the operator was targeting French media and entertainment websites and the education sector. Google does not name which web applications it was analyzing.
  • The PaperCut case: in late August 2026, two flaws in PaperCut print management servers were exploited. Per the report, security researchers and the company Huntress had working exploits within a day of disclosure using LLMs, and GreyNoise reported a campaign four days after disclosure against hundreds of organizations in which the attacker "likely" used LLMs.
  • PHP: Google says the PHP flaws patched on 21 October 2026 (CVE-2026-109618 and CVE-2026-109619) were discovered using LLM-enabled analysis, and that hundreds of non-functional proof-of-concept exploits, likely AI-generated, followed the patch.
  • The numbers: according to a September GTIG analysis, overall disclosures and exploits nearly doubled between January and August 2026, while exploited zero-days went from a 2025 average of 8 per month to 11 per month in 2026, spiking to 22 in August.
  • Stolen AI keys: GTIG describes "LLMjacking", stealing credentials for AI services and reselling access. In Mandiant incident response cases, fraudulent charges exceeded USD 100,000 in a single day and, in the worst cases, reached USD 1 million cumulatively. A common entry point was valid credentials left in public GitHub repositories.

What changes and what doesn't

  • The window shrinks. If a system can read a patch and have an exploit ready in hours, the gap between "update released" and "I'm being attacked" gets shorter. Leaving a plugin or module update for weeks carries more risk than it used to.
  • The defense is the same. The report does not describe a new class of flaw: what speeds up is exploitation of flaws that already have fixes. Against an up-to-date site, this method has no already-fixed flaw to exploit.
  • They are noisy. GTIG says these frameworks currently "trade stealth for velocity": they generate a lot of traffic and leave recognizable traces, so a web application firewall or reviewed logs can catch them.
  • What the report doesn't say: it gives no count of websites attacked this way and, while one figure shows an agent framework inspecting a WordPress plugin, it does not say a flaw was found in it. The zero-day figures are GTIG's own, and the PaperCut and LLMjacking cases come from third parties or Google's investigations.

What to do on your site

  1. Turn on automatic security updates. WordPress installs minor core releases automatically by default; also enable auto-updates for plugins that update cleanly. On Drupal, subscribe to drupal.org security advisories and set yourself a 48-hour deadline to apply them.
  2. List what you have installed. Plugins, modules, themes and your PHP version. Delete what you don't use: a deactivated plugin is still on the server and can still be attacked.
  3. Check your PHP version. In your hosting panel, or under Site Health in WordPress (Status report in Drupal). If you are not on the release with the 21 October fix, ask your host when they will apply it.
  4. Put a web application firewall in front. A WAF from your host or a service like Cloudflare can block attacks on known flaws while you update, and these systems make a lot of noise.
  5. Protect your AI keys. If your site or automations call OpenAI, Google, Anthropic or other model APIs: keep keys out of code and public repositories, set a spending limit in the provider's dashboard and check daily usage. An unexpected spike may mean a stolen key.
  6. Check GitHub. If you ever pushed your site's code to a public repository, look for .env files, passwords or tokens. Treat anything that was exposed as stolen: issue a new key and revoke the old one.

Related: FBI Warns of ARTEX AI Attacks: What to Check on Your Site

Sources

Updates: any changes backed by a primary source will be added here.