Google Warns AI-Powered Cyber Attack Frameworks Are Accelerating Intrusions

Published:

Google has reported a growing shift in how cyber threat actors use artificial intelligence, with attackers moving beyond manually prompting large language models (LLMs) towards deploying autonomous, multi-agent frameworks to conduct different stages of cyber intrusions. According to Google Threat Intelligence Group (GTIG), open-source offensive platforms such as ARTEX, CAIRN, CyberStrikeAI and PentAGI are increasingly being incorporated into attack operations, allowing threat actors to automate reconnaissance, vulnerability scanning, lateral movement and other activities. In its October 30, 2026 report, Google said it had observed several of these frameworks being used during incident response investigations conducted by Mandiant. The systems coordinate multiple agents through open-source and commercial AI models, enabling them to interpret operational objectives, select tools, respond to errors and pursue different attack paths with limited human intervention. Google described the development as a significant change in the pace and scale of cyber operations, as automated frameworks can execute sequences of activities in seconds or minutes that would otherwise require considerably more manual effort. The underlying techniques, including exploiting unpatched systems and using compromised credentials, are not entirely new. However, the ability to automate and connect these activities allows attackers to pursue more targets and carry out multiple stages of an intrusion at greater speed. GTIG said financially motivated cybercriminals and suspected cyber espionage operators have both been observed experimenting with these capabilities, indicating that agentic AI is becoming part of a broader range of threat activity.

Another area of concern involves the use of LLMs to analyse software updates and accelerate vulnerability exploitation. Google said threat actors are using AI-assisted workflows to compare software versions, examine security patches and identify changes that could reveal how a vulnerability works. By analysing source code before and after a security update, attackers can potentially identify the weakness addressed by a patch and develop an exploit without having to discover the original vulnerability independently. The approach can reduce the time between public disclosure of a security issue and attempts to exploit systems that have not yet been updated. Google cited activity involving vulnerabilities in PaperCut NG/MF as an example of AI-assisted analysis and exploitation development. Such workflows can draw on publicly available source code, release notes, security research and proof-of-concept material to automate tasks that previously required specialist research. GTIG also highlighted the use of services such as ZeroHollow, an inference platform associated with exploit development and LLM routing, alongside Agent Zero, an open-source agent framework offering autonomous code execution and dynamic tool generation. These developments show how AI capabilities are being integrated into existing offensive security tools and criminal workflows. However, the use of an individual framework or AI service does not by itself establish malicious activity, as similar technologies can also support legitimate penetration testing and security research. Google’s findings concern their observed use in unauthorised operations and the potential for these capabilities to increase the speed of vulnerability research and exploitation.

Google also reported that threat actors are targeting AI infrastructure and cloud resources to obtain access to computing capabilities and generate revenue. One emerging activity is LLMJacking, in which attackers steal or misuse credentials and cloud access to operate large language models through compromised enterprise environments. Rather than paying for their own access, operators use stolen API keys, identity tokens or cloud credentials to access AI services and resell that capacity through underground proxy platforms known as LLM mixers. Google has identified services including APIMart and JetNight in its reporting on this market. These platforms can provide customers with access to AI models at reduced prices while the underlying usage costs are charged to compromised organisations. The activity creates financial risks for businesses and can expose cloud environments to further unauthorised access. In a separate development, GTIG identified the use of AI capabilities within phishing-as-a-service operations. One example is Jieyou, a Chinese-language phishing platform that incorporates an LLM chatbot to help customers configure phishing panels and modify existing templates through natural-language instructions. Google said this approach allows affiliates with limited technical knowledge to customise phishing pages and social engineering messages without manually editing the underlying material. The chatbot can help users adapt lures and interfaces for different scenarios, making the service easier to operate and potentially expanding the number of people capable of conducting phishing campaigns. GTIG described the development as an example of AI functioning as a modular tool within cybercrime services rather than simply generating isolated malicious messages.

The findings point to a broader expansion of AI-assisted activity across the cyber threat landscape, from vulnerability research and automated intrusion operations to credential theft, cloud resource abuse and phishing. Google said the growing availability of open-source agentic frameworks has lowered the technical barriers to combining different offensive capabilities, while allowing operators to automate tasks across several stages of an attack. The trend also creates challenges for organisations that rely on traditional security controls and response processes, particularly when attackers can move quickly between compromised accounts, exposed services and internal systems. Security teams therefore need to maintain timely software updates, monitor unusual authentication and cloud activity, protect API keys and other credentials, and restrict access to sensitive infrastructure. Organisations using AI services should also review permissions, monitor consumption patterns and investigate unexpected increases in model usage that could indicate compromised access. The distinction between legitimate security testing and unauthorised activity remains important because many of the frameworks mentioned by Google are openly available and can serve legitimate testing purposes. The risk arises when these capabilities are used against systems without authorisation or combined with stolen credentials and exploited vulnerabilities. GTIG’s observations indicate that agentic AI has moved beyond limited experimentation in some threat operations, although fully autonomous and highly sophisticated attacks remain a developing possibility rather than a universal characteristic of current incidents. The findings underline the need for defenders to adapt monitoring and incident response practices as attackers increasingly use AI to accelerate established cyber techniques.

Source

Follow the SPIN IDG WhatsApp Channel for updates across the Smart Pakistan Insights Network covering all of Pakistan’s technology ecosystem.

Related articles

spot_img