Skip to content

Independent AI & deep-tech news

Artificial Intelligence

Google Warns Hackers Are Using AI Agents to Automate Cyberattacks

Google Warns Hackers Are Using AI Agents to Automate Cyberattacks
Image Credit: https://opendatascience.com/
On this page
  1. AI Agents Are Accelerating Cyberattacks
  2. Hackers Use AI to Reverse-Engineer Software Patches
  3. LLMjacking Creates Additional Security Risks
  4. What Does This Mean for AI Security Teams?

How are hackers using AI agents to release cyberattacks faster? As per to a new report from Google’s Threat Intelligence Group (GTIG), threat actors are deploying autonomous AI frameworks to discover vulnerabilities, reverse-engineer software patches, and automate exploitation. These capabilities permit attackers to execute multi-stage intrusions with less human intervention, clearly decreasing the time organizations have to respond to developing threats.

The November 16 report from Google emphasize a growing shift from manually prompting large language models (LLMs) closer to coordinated, agent-driven cyberattacks.

AI Agents Are Accelerating Cyberattacks

Google diagnosed numerous open-source, multi-agent penetration-testing out frameworks that attackers have adopted or experimented with, including ARTEX, CAIRN, CyberStrikeAI, and PentAGI.

These frameworks coordinate specialised AI agents to carry out tasks along with vulnerability discovery, reconnaissance, exploit development, and attack execution. Unlike conventional automation tools, agentic frameworks can examine execution feedback and ajust their approach without demanding continuous human direction.

GTIG observed financially encouraged attackers and suspected Chinese espionage groups experimenting with these capabilities.

The findings suggest that attackers can scale existing techniques more efficiently instead of growing wholly new exploitation techniques.

Hackers Use AI to Reverse-Engineer Software Patches

One of Google’s most considerable findings includes the automated analysis of software updates.

Attackers are using LLM-powered pipelines to compare patched and vulnerable code versions through differential analysis. This technique supports detect security flaws that developers lately addressed.

Once attackers detect a vulnerability, AI tools can boost up evidence-of-concept development and potential exploitation. Google emphasized activity including PaperCut NG/MF vulnerabilities disclosed in September 2026 for example of this emerging method.

The implications extend beyond individual vulnerabilities. Organizations traditionally depend on a window among vulnerability disclosure and widespread exploitation to deploy security updates.

AI-assisted exploitation threatens to compress that window, making rapid patch management increasingly vital.

LLMjacking Creates Additional Security Risks

Google also documented the developing misuse of enterprise AI infrastructure through assaults known as LLMjacking.

In those incidents, attackers steal credentials or compromise cloud resources to gain unauthorized access to commercial AI models. Instead of using of the stolen resources solely for their own operations, some attackers resell access through underground proxy services.

Google detected services such as APIMart and JetNight as examples of these so-called LLM mixers. Meanwhile, AI capabilities are attaining phishing-as-a-service platforms. The Chinese-language Jieyou platform incorporates an LLM assistant that permits to customers phishing pages through natural language commands.

These developments lower the technical demands for releasing certain cyberattacks while expanding their potential reach.

What Does This Mean for AI Security Teams?

For AI engineers and cybersecurity professionals, Google’s findings emphasize the need to reassess traditional defensive timelines.

Security teams should prioritize fast vulnerability remediation, persisted tracking, credential protection, and automated detection. Organizations deploying AI systems must also monitor API credentials, cloud permissions, and unusual model usage to detect potential resource abuse.

The wider challenge includes defending against adversaries that increasingly use the same automation capabilities available to legitimate developers.

Written by

Tarun Khanna

579 articles on DeepTechBytes