Robots Are Better Than People home
Subscribe
Silhouetted analyst in a dark security operations center watching code diffs and network graphs

Google Finds AI Agents Turning Security Patches Into Live Attacks

Google's threat intelligence team says it caught a multi-agent AI system doing every step of an attack on its own, from reading a fresh patch to running the exploit. Here is what was found, what was not, and what defenders should change this week.

For years the security industry has warned that someone would wire a large language model into the whole attack chain and walk away. On November 16, 2026, Google Threat Intelligence Group (GTIG) and Mandiant said they had found exactly that. In their quarterly AI threat report, the team wrote that "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 sentence is the news. The rest of this piece explains what the system did, what Google did not say, and why the people who should care most are the ones running ordinary websites.

What Google Actually Found

GTIG says it located an exposed staging server, left reachable on the internet by its operator. On it sat a multi-agent framework the researchers call exploit_agent. According to the report, it "integrates with multiple frontier and local LLMs" and is set up for "continuous differential analysis of popular web application software."

In plain terms, the system watches for software updates. When a vendor ships a fix, the agents compare the code before and after the patch, work out which weakness the patch closes, draft an exploit for that weakness, and then, in Google's words, "automatically executes them without human intervention."

The server held more than code. GTIG says it also found data taken from victims and scripts for post-compromise work: internal reconnaissance, brute-forcing passwords, installing webshells, and even removing other attackers' webshells. Victim telemetry pointed to French websites in media and entertainment, plus the education sector. The configuration files carried Russian-language comments, one of which labels the stored keys as "LLM API keys for auto-exploitation (exploit_agent pilot)."

Field note: Google did not name a group, a country or a motive, and it did not say how many sites were hit. Russian-language comments are a clue about the operator, not an attribution. Several outlets wrote "Russian hackers"; the report itself is more careful, and so are we.

Why the Patch Itself Became the Weapon

Every security update is also a map. It shows, line by line, where the old code was weak. Skilled humans have always been able to read that map; the industry calls the result an n-day exploit, because the flaw is public but many systems are still unpatched.

GTIG's argument is that language models make this kind of work cheap. The report notes attackers "can focus research on a finite set of source materials," such as source code versions, release notes and published proofs of concept, while hunting brand-new zero-days "requires casting a wider net." Automating the first job shrinks what defenders call the patch window: the days or weeks between a fix being released and a fix being installed.

Google's own numbers show the pressure building. The team says overall vulnerability disclosures and exploits "nearly doubled between January and August 2026," while zero-day exploits rose from a 2025 average of 8 per month to 11 per month in 2026, before spiking to 22 in August.

StepHuman-led n-day attackWhat GTIG observed
Notice a patchAnalyst reads advisoriesContinuous automated monitoring
Find the flawManual diffing, hours to daysLLM agents compare versions
Write the exploitSpecialist timeAgents draft it
LaunchOperator decidesExecuted "without human intervention"

The Wider Pattern: Agent Frameworks in Real Intrusions

exploit_agent was not the only finding. GTIG says it has watched threat actors adopt open source "agentic penetration-testing frameworks," tools originally built for legitimate security testing, during live intrusions. It names ARTEX, CAIRN, CyberStrikeAI and PentAGI, and says it saw CAIRN used against targets including the US aerospace sector, and a customized PentAGI platform used for reconnaissance and vulnerability scanning against organizations in Asia between late August and early October 2026.

The report also describes LLMJacking, where criminals steal credentials for AI services and resell access, and influence operations that use models to run synthetic personas at scale. Those are separate stories; we will cover them on their own.

The Messy Middle: Fast, but Not Quiet

There is a piece of good news buried in the report, and defenders should take it seriously. GTIG writes that these frameworks "currently trade stealth for velocity." Unrestricted autonomous execution "generates forensic traffic, relies on recognizable identifiers, and leaves behind distinct code-level traces." Translation: the robots are fast and loud. Organizations that actually watch their logs have a real chance to catch them.

The same paragraph carries the warning. Google expects "not only reduced human involvement in these operations but also increased stealth and sophistication of the activity" as the underlying models improve. The noisy phase will not last.

The lower barrier also cuts the other way. GTIG notes a wave of "LLM-generated exploits that are claimed to work but don't," which creates extra verification work for defenders who have to sort real threats from junk.

Decision Matrix for Site Owners

  • If you run WordPress or another popular CMS: turn on automatic minor and security updates, and remove plugins you do not use. Each one is another patch to race.
  • If you cannot patch within days: put a managed web application firewall in front of the site and ask your host about virtual patching.
  • If you run a security team: measure your mean time to patch for internet-facing apps. Google's point is that this number now matters more than almost anything else.
  • If you use AI APIs: store keys in a secrets manager, cap spend, and alert on unusual usage. LLMJacking turns your account into someone else's attack budget.

Our Read: The Inflection Point

We are not going to repeat lines that are not in the report. Some social posts quoted Google declaring that "the era of autonomous attacks has arrived." We searched the published report and could not find that sentence. What Google did write is strong enough: it calls the moment "a critical inflection point" and describes "a paradigm shift towards intrusions occurring at machine speed."

Here is why that matters. Until now, the slow part of turning a patch into an attack was a skilled person's time. That constraint is gone for at least one operator. The mechanics are old (unpatched software, stolen passwords), but the clock has changed. Defenders who still patch monthly are now racing software that never sleeps, and the only fair fight is automation on the defense side too: automatic updates, managed firewalls, and alerting that someone actually reads.

The bottom line: A real attacker used AI agents to go from published patch to executed exploit with no human step. The tools are noisy today. Your patch speed is now your most important security number.

Sources

We checked every fact, figure and quote in this story against the sources below. Quotes are reproduced as published. Images are AI-generated illustrations, not photos of the events described.

  1. Google Threat Intelligence Group: GTIG AI Threat Tracker, Agentic Threat Activity Crosses a Significant Threshold (Nov 16, 2026)
  2. The Register: Google says AI agent turned a security patch into a working exploit with no human help
  3. CyberScoop: Google says Russian-speaking hackers ran the first fully automated AI cyberattack
  4. TechTarget: Google finds AI agents autonomously turning patches into exploits
  5. CISA Known Exploited Vulnerabilities Catalog

RABTP Newsroom

Every story is researched against primary sources, then reviewed by publisher Sean Hakes before it goes live. We cover threats from the defender's side and never publish exploit details. Read our standards or send a correction.

Keep reading

The White-Hat Brief

Weekly: what the machines learned, and how to stay ahead

One email a week. Verified stories, plain-language defenses, no hype. Unsubscribe anytime.