Google says attackers now use AI agents to find and exploit software flaws

Editorial illustration: A floating robot uses articulated tools and an amber scanning beam to examine a glowing red gap between layers of a circuit-like structure.

In brief

  • Google's GTIG says attackers used LLM frameworks to automate vulnerability discovery and exploitation, a first.
  • An exposed staging server hosted a multi-agent framework that compared code before and after patches, GTIG says.
  • Exploited zero-days averaged 11 per month in 2026 through August, reaching 22 in August, per Google's report.
  • Automated workflows leave organizations less time to apply security patches, Google warns.

What GTIG found on the staging server

The evidence came from a tool found on an exposed staging server, according to the GTIG report as reported by Cointelegraph. It compared software versions, identified potential flaws, and then generated and executed exploit code.

No step-by-step human direction, per GTIG.

The server hosted a multi-agent vulnerability research and weaponization framework that ran continuous differential analysis (comparing code before and after patch announcements on popular web applications). GTIG said attackers were using AI agent frameworks to automate patch reverse engineering this way and to weaponize N-day vulnerabilities, which are publicly known security flaws.

"In the past quarter, GTIG identified threat actors using LLM frameworks to fully automate vulnerability discovery and exploitation without a human in the loop for the first time"

That's a change from GTIG's September report, which said it hadn't yet observed threat actors deploying fully autonomous pipelines against targets in the wild.

A shorter window to patch

According to Google, vulnerability disclosures more than doubled between January and August. Exploited zero-day vulnerabilities averaged eight per month in 2025 and 11 per month from January through August 2026, the report said, reaching 22 in August alone.

The report cited an August 2026 PaperCut campaign in which attackers gained domain administrator-level control within minutes, according to cybersecurity firm GreyNoise. Cloudflare separately reported attempts to exploit PHP flaws against its customers in October. GTIG said the two PHP vulnerabilities involved were discovered using LLM analysis, after technical information was published before the Oct. 21 patch.

Google's assessment is that LLMs are letting attackers scale exploitation pipelines against internet-facing servers without human intervention. It's a warning about the patch window, and the company framed it that way: automated workflows give defenders less time to fix flaws before exploits exist.

Anthropic's separate warning

Google isn't the only AI developer raising this. In a September report, Anthropic warned that AI-assisted attacks had evolved into multi-agent operations automating most of the cyberattack cycle, from discovery to exploitation and data theft. Those findings are Anthropic's own and were published before GTIG's Monday report.

Frequently asked questions

What did Google's threat intelligence team find?

Google Threat Intelligence Group said it had, for the first time, identified attackers using AI large language model frameworks to automate vulnerability discovery and exploitation without step-by-step human direction. The evidence included a tool on an exposed staging server that compared software versions, identified flaws, and generated and executed exploit code.

How fast are zero-day vulnerabilities being exploited in 2026?

According to Google, exploited zero-day vulnerabilities averaged eight per month in 2025 and 11 per month from January through August 2026, reaching 22 in August. Vulnerability disclosures more than doubled between January and August.

Why does Google say AI-driven exploitation matters for patching?

Google warned that automated workflows leave organizations less time to apply security patches before attackers develop exploits. GTIG said attackers were using AI agent frameworks to reverse engineer patches through differential analysis and weaponize publicly known N-day vulnerabilities.