Autonomous AI cyberattack: what Google's AI Threat Tracker found, and why "zero humans" is hype

Cybersecurity

Quick answer: Google’s Threat Intelligence Group (GTIG) says it found an AI framework, exploit_agent, on an attacker’s exposed server. It compares web software before and after a patch, finds the bug, writes an exploit and runs it with no person involved. That part is real. A viral graphic’s “0 humans, at any stage” is not what Google’s own chart says. Patch internet-facing software fast. (Google GTIG, Nov 16)

Affected: Websites running “popular web application software,” per Google, which does not name a full product list or a CVE. One screenshot in Google’s report shows the framework inspecting the Spectra (Ultimate Addons for Gutenberg) WordPress plugin, version 2.19.25. Victim data pointed at French media, entertainment and education websites. Fix: No single patch. Install vendor security updates on anything that faces the internet as soon as they ship, and follow Mandiant’s hardening guide (Google/Mandiant, Oct 30). Exploited yet? Yes. Google says the attacker’s server held “data exfiltrated from targeted victims,” per Google Threat Intelligence Group, Nov 16.

What happened

On Nov 16, 2026, Google published its latest AI Threat Tracker. The headline finding, in Google’s words: “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 “first” is Google’s claim. Nobody outside Google has confirmed it yet. It is also not the first “first.” On Nov 13, 2025, Anthropic said attackers had used its Claude Code tool against “roughly thirty global targets” and wrote: “We believe this is the first documented case of a large-scale cyberattack executed without substantial human intervention.” In Anthropic’s case, people still stepped in at “perhaps 4-6 critical decision points per hacking campaign.” Google’s claim is narrower: no person at all, but only for finding and exploiting the bug. (Anthropic, Nov 13, 2025)

Here is what Google says it found. GTIG “identified an exposed staging server hosting a multi-agent vulnerability research and weaponization framework, dubbed exploit_agent, that integrates with multiple frontier and local LLMs.” Google describes the loop this way: “It compares code before and after patch announcements, analyzes findings to identify potential vulnerabilities, drafts exploits, and automatically executes them without human intervention.”

In plain terms, every security update tells you where the old bug was. Compare the old code with the new code and you can work backward to the hole. People have done this by hand for years. Google says this tool does it automatically, then attacks sites that haven’t installed the update yet.

The same server held stolen data and scripts for what attackers do once inside: “internal reconnaissance, brute-forcing, installing webshells, and removing other attackers’ webshells.” Google says “Victim telemetry on the server suggests the operator was specifically interested in targeting French websites associated with the media and entertainment industries, as well as the education sector.” A config file on the server had Russian-language comments. Google does not name the attacker, the victims or how many sites were hit.

Where the hype comes in

A graphic going around on X says “The first fully autonomous AI cyberattack in < 3 hrs,” with tiles for “0 Humans in the loop, at any stage” and “150+ Attack tools, one AI cyberattack framework.” It pulls from three different tools and one chart that says something else.

One more thing to watch for. Some posts about the report quote Google as saying the attackers’ methods have not changed. Google’s report does not say that in those words. What it says is: “While the mechanics of the intrusions remain the same, such as using compromised credentials and targeting unpatched systems, the speed and scale at which frameworks can exploit these known weaknesses introduces a new defensive challenge.”

The good news from the same report: these tools are noisy. Google says they “currently trade stealth for velocity,” and their unrestricted runs leave “forensic traffic,” “recognizable identifiers” and “distinct code-level traces.”

We covered a related case earlier: an unreleased Claude model writing exploit chains, and why you should update Chrome to 156.

Am I affected?

Google has not published a product list, a CVE or a way to check a specific site for exploit_agent. Ask two questions instead.

  1. Do you run a website or web app on software someone else patches, like WordPress and its plugins? If a security update came out and you haven’t installed it, Google’s point is that the gap between the update and an attack is getting shorter. Google says AI is “shrinking the patch window for organizations.”
  2. Do you pay for AI services with API keys? Google’s report also covers LLMJacking, where attackers steal cloud or AI keys and run up the bill. Mandiant handled cases where “single-day fraudulent charges exceeded $100,000 USD,” and “In the most severe cases, cumulative resource abuse reached $1 million USD before remediation.”

What to do now

  1. Patch internet-facing systems first, and fast. Websites, plugins, VPNs, firewalls and remote-access tools come before office PCs. For WordPress, turn on automatic updates for plugins you trust, and remove plugins you don’t use.
  2. Use short-lived credentials. Mandiant’s hardening guide recommends “replacing long-lived secrets with short-lived and tightly scoped credentials.” Rotate any API key or password that has sat unchanged for a long time. (Google/Mandiant, Oct 30)
  3. Use phishing-resistant sign-in for admin accounts. The same guide says to pair admin access controls “with phishing-resistant authentication such as FIDO2 or passkeys where supported.” Start with your website, hosting and cloud admin logins.
  4. Set budget alerts on AI and cloud accounts. Google’s LLMJacking guidance: “Implement granular API quotas on LLM services to limit request volume per model. Utilize spending caps to set budgets and avoid over-spending.” It also recommends automated secret scanning so keys don’t sit in code. (Google GTIG, Nov 2)

Sources

AltTab Dad