Google Finds AI Agents Turning Software Patches Into Real-World Attacks
Google reports a shift toward automated exploitation in live cyber intrusions. Photo: Sasun Bughdaryan / Unsplash
Cybersecurity & Privacy

Google Finds AI Agents Turning Software Patches Into Real-World Attacks

Google’s investigators found automated exploitation in live intrusions, raising the stakes for defenders racing to patch exposed systems.

By Marcus Lee • 4 mins read Edited by Maria Konash Published: Updated:

Key Notes

  • Google observed an AI workflow that compared software patches, created exploits and executed them without continuous human direction.
  • The report also describes poisoned coding-assistant settings and development of an influence platform designed for more than 100,000 accounts.
  • A separate Mandiant investigation found stolen cloud and model access generating cumulative fraudulent charges of up to $1 million.

Google says attackers are now using AI workflows to inspect software updates and turn the fixes into working attacks without continuous human direction. Its November 16 threat update describes a shift from chatbots assisting an operator to systems carrying out substantial parts of an intrusion.

In a companion announcement, the company said automated frameworks can repeatedly identify what a patch repaired and build an attack around it. The finding concerns criminal use of AI against real targets; it does not mean Google’s own systems were the victims.

Automated Exploitation Narrows the Patch Window

In the full report, Google Threat Intelligence Group identified a framework called exploit_agent on an exposed server. It compared pre- and post-patch code, produced exploits and executed them automatically. Investigators also found stolen victim data.

This is Google’s first reported observation of that fully automated vulnerability-research and exploitation workflow. The report also describes tools that automate selected phases of intrusions. The degree of autonomy varies by case, and human operators still supply goals, infrastructure or access.

A software fix changes what both defenders and attackers can inspect. Organizations that have not yet installed it remain exposed while others analyze the difference. The practical risk is a shorter interval between a repair becoming public and vulnerable systems being attacked.

Recruiter Lures Reach Coding-Assistant Settings

The report also describes suspected North Korea-linked operators targeting cryptocurrency businesses in North America and Europe. After recruiter-themed lures and initial compromise, attackers altered internal repository settings for Anthropic’s Claude Code so launching the assistant downloaded the SOMBERMEME backdoor.

The sequence matters: the assistant’s configuration became a way to extend an existing intrusion. A familiar development tool can execute dangerous actions when the project material it trusts has been changed. Reviewing code alone may miss settings that influence what happens when the tool starts.

Stolen AI Access Creates Million-Dollar Bills

A separate November 2 Mandiant investigation documented attackers stealing access to cloud and model services, then consuming or reselling the resources. Fraudulent charges exceeded $100,000 in a single day in some incidents, with cumulative charges at some organizations reaching $1 million before access was revoked.

The activity, known as LLMJacking, turns model credentials into a resource that criminals can monetize. Unauthorized consumption can be costly even when attackers are not trying to steal the organization’s own model or training data. The reported totals are charges from observed incidents, rather than a standard cost for every compromise.

Our reporting on recent GitHub compromises also examined stolen cloud and AI keys. Model access now belongs inside the same security and spending controls as other privileged infrastructure.

Influence Tools Are Being Built for Scale

Separately, Google found development of an influence platform designed to manage more than 100,000 social accounts. That is a planned capacity, rather than confirmation that such a network was already operating.

The distinction separates the capabilities being engineered from the activity investigators could demonstrate. Automating the creation and management of online personas could lower the cost of coordinated manipulation, but a design target alone cannot establish how many accounts were deployed or how many people they reached.

Speed Still Leaves Evidence

Google and Mandiant’s earlier October 30 investigation described agentic tools conducting rapid, noisy trial-and-error during live intrusions. In one example, activity generated hundreds of log events. This creates opportunities for detection even while the pace makes a manual response harder.

That investigation recommended short-lived credentials, limited privileges, network segmentation and behavioral detection. These controls reduce what an automated workflow can reach after gaining access. Reviewing coding-tool settings and watching model usage can address other parts of the same problem.

Researchers expect future frameworks to require less human intervention and leave fewer traces. That is their assessment of the direction of the threat, not a claim that today’s systems already conduct every stage of every attack independently.

Disclaimer: AIstify is an independent media brand owned and operated by NuvexMedia LLC, publishing news, research, and insights on artificial intelligence, emerging technologies, automation, and related industries. NuvexMedia LLC invests in and collaborates with companies across the AI, technology, software, and digital innovation sectors. These relationships do not influence AIstify’s editorial coverage, and the publication maintains full editorial independence to provide accurate, timely, and objective information. © 2026 NuvexMedia LLC. All rights reserved. This content is for informational purposes only and should not be considered legal, tax, investment, financial, or other professional advice.

AI & Machine Learning, Cybersecurity & Privacy, News