Image Source: Google Cloud
Cyber attacks are entering a faster, more automated era, according to a new report from the Google Threat Intelligence Group (GTIG) and Mandiant. Published November 16, 2026, the assessment says criminals, espionage groups, and influence operators are using generative AI to compress attack timelines, expand their reach, and complete complex operations with less human involvement.
The report describes a major shift from basic prompting to autonomous, multi-stage workflows. Threat actors are now experimenting with agentic frameworks that can perform reconnaissance, identify weaknesses, generate payloads, and adjust their actions based on real-time results.
Cyber Attacks Move Toward Machine-Speed Operations
GTIG observed open-source offensive frameworks including CAIRN, CyberStrikeAI, HexStrike AI, and PentAGI being used or tested in real-world activity. These systems connect large language models with security tools and specialized agents, allowing attackers to automate portions of a digital intrusion.
CAIRN uses an observe, orient, decide, and act loop to coordinate multiple agents. CyberStrikeAI combines a command-and-control backend with more than 100 security tools, while HexStrike AI connects language-model clients to more than 150 offensive utilities. PentAGI has also been adapted for systematic reconnaissance and vulnerability scanning against organizations in Asia.
Google said the technology currently favors speed over stealth. Automated operations can generate unusual forensic activity, recognizable code traces, and other indicators that help defenders detect them. However, GTIG expects future systems to become quieter, more capable, and increasingly independent.
Explosive Automation of Vulnerability Exploitation
One of the most significant findings involves the automation of vulnerability research. Threat actors are using AI agents to compare software versions, study security patches, analyze release notes, draft proof-of-concept code, and test exploits.
GTIG identified an exposed platform called exploit_agent that was designed to conduct continuous differential analysis of web applications. The system could identify potential flaws and automatically attempt exploitation without a human operator. Investigators also found scripts for internal reconnaissance, brute-force activity, webshell installation, and post-compromise operations.
The report noted that vulnerability disclosures and exploitation activity nearly doubled between January and August 2026. Although zero-day activity rose only modestly for much of the year, it reached 22 incidents in August, compared with an average of eight per month in 2025.
Researchers also highlighted attacks involving PaperCut servers and PHP software. In some cases, working exploits were reportedly developed within a day of disclosure. The rapid appearance of hundreds of fake or ineffective proof-of-concept claims also created additional verification work for security teams.
LLMJacking Turns Stolen Cloud Access Into Profit
Financially motivated groups are not only using AI to launch attacks; they are also stealing access to AI services and cloud computing resources. This activity, known as LLMJacking, can cause enormous unexpected bills for victims.
Mandiant reported incidents in which fraudulent charges exceeded $100,000 in a single day. In the most serious cases, unauthorized use accumulated to approximately $1 million before the activity was stopped.
Attackers commonly steal credentials, hijack active sessions, create unauthorized service accounts, or attach administrative permissions to existing cloud identities. They may then activate language-model services, request GPU capacity, or deploy automated tools against other targets.
Underground marketplaces known as LLM Mixers are making the problem worse. These services resell access to stolen accounts, model quotas, and compromised computing infrastructure. Advertised prices ranged from $0.10 to $25 per million tokens, depending on the model and account type.
AI-Powered Influence Operations Expand
GTIG also said information operations have moved beyond experimentation. State-backed groups, commercial Influence-as-a-Service vendors, and inauthentic political monetization networks are using AI to produce synthetic media, create localized personas, and distribute content across multiple platforms.
Some campaigns use regional slang, deliberate spelling mistakes, artificial camera flaws, residential proxies, and aged social media accounts to appear more authentic. Other platforms are designed to manage thousands of personas, virtual smartphones, disposable email accounts, and automated posting schedules.
Not every campaign is politically motivated. Inauthentic political monetization networks use AI-generated content to exploit polarization, attract clicks, and increase advertising revenue. Their output can resemble state-sponsored propaganda, making attribution more difficult.
Defenders Face a Critical New Security Challenge
Government-backed groups are also adopting AI. GTIG observed PRC-nexus actors using language models for reconnaissance, target research, exploit development, and analysis of stolen data. DPRK-linked operators reportedly used generative tools to prepare job applications and interviews while posing as remote technology workers.
In one reported supply-chain attack, a suspected DPRK-linked cluster manipulated configuration files used by an autonomous coding assistant. The changes were designed to trigger the download of a backdoor when developers launched the tool.
Google said its defenses include model safeguards, threat intelligence, automated monitoring, account disruption, red-team testing, and enterprise tools such as Google AI Threat Defense. The company also said insights from investigations are fed back into safety classifiers and platform protections.
For organizations, the report reinforces several urgent priorities: patch internet-facing systems quickly, protect cloud and AI credentials, monitor unexpected model usage, review service-account permissions, and examine coding-assistant configurations for unauthorized changes. As AI-powered cyber attacks become faster and more accessible, traditional security controls must operate at machine speed as well.
Frequently Asked Questions
What are agentic cyber attacks?
Agentic cyber attacks use autonomous or semi-autonomous AI systems to perform multiple stages of an intrusion, such as reconnaissance, vulnerability discovery, exploitation, and post-compromise activity.
What is LLMJacking?
LLMJacking is the unauthorized use of stolen cloud accounts, AI credentials, model quotas, or computing resources to run language models and generate profit for attackers.
Which AI frameworks are being used by threat actors?
GTIG identified activity involving CAIRN, CyberStrikeAI, HexStrike AI, PentAGI, ARTEX, and custom-built orchestration systems.
How can companies reduce these risks?
Organizations should use strong identity controls, multifactor authentication, least-privilege permissions, rapid patching, cloud-cost alerts, continuous monitoring, and strict oversight of AI coding and automation tools.